… | |
… | |
8 | |
8 | |
9 | perl -MFaster ... |
9 | perl -MFaster ... |
10 | |
10 | |
11 | =head1 DESCRIPTION |
11 | =head1 DESCRIPTION |
12 | |
12 | |
13 | This module implements a very simple-minded JIT. It works by more or less |
13 | This module implements a very simple-minded "JIT" (or actually AIT, ahead |
14 | translating every function it sees into a C program, compiling it and then |
14 | of time compiler). It works by more or less translating every function it |
15 | replacing the function by the compiled code. |
15 | sees into a C program, compiling it and then replacing the function by the |
|
|
16 | compiled code. |
16 | |
17 | |
17 | As a result, startup times are immense, as every function might lead to a |
18 | As a result, startup times are immense, as every function might lead to a |
18 | full-blown compilation. |
19 | full-blown compilation. |
19 | |
20 | |
20 | The speed improvements are also not great, you can expect 20% or so on |
21 | The speed improvements are also not great, you can expect 20% or so on |
21 | average, for code that runs very often. |
22 | average, for code that runs very often. The reason for this is that data |
|
|
23 | handling is mostly being done by the same old code, it just gets called |
|
|
24 | a bit faster. Regexes and string operations won't get faster. Airhtmetic |
|
|
25 | doresn't become any faster. Just the operands and other stuff is put on |
|
|
26 | the stack faster, and the opcodes themselves have a bit less overhead. |
22 | |
27 | |
23 | Faster is in the early stages of development. Due to its design its |
28 | Faster is in the early stages of development. Due to its design its |
24 | relatively safe to use (it will either work or simply slowdown the program |
29 | relatively safe to use (it will either work or simply slowdown the program |
25 | immensely, but rarely cause bugs). |
30 | immensely, but rarely cause bugs). |
26 | |
31 | |
|
|
32 | More intelligent algorithms (loop optimisation, type inference) could |
|
|
33 | improve that easily, but requires a much more elaborate presentation and |
|
|
34 | optimiser than what is in place. There are no plans to improve Faster in |
|
|
35 | this way, yet, but it would provide a reasonably good place to start. |
|
|
36 | |
27 | Usage is very easy, just C<use Faster> and every function called from then |
37 | Usage is very easy, just C<use Faster> and every function called from then |
28 | on will be compiled. |
38 | on will be compiled. |
29 | |
39 | |
30 | Right now, Faster will leave lots of F<*.c>, F<*.o> and F<*.so> files in |
40 | Right now, Faster can leave lots of F<*.c> and F<*.so> files in your |
31 | F</tmp>, and it will even create those temporary files in an insecure |
41 | F<$FASTER_CACHEDIR> (by default F<$HOME/.perl-faster-cache>), and it will |
32 | manner, so watch out. |
42 | even create those temporary files in an insecure manner, so watch out. |
33 | |
43 | |
34 | =over 4 |
44 | =over 4 |
35 | |
45 | |
36 | =cut |
46 | =cut |
37 | |
47 | |
38 | package Faster; |
48 | package Faster; |
|
|
49 | |
|
|
50 | no warnings; |
39 | |
51 | |
40 | use strict; |
52 | use strict; |
41 | use Config; |
53 | use Config; |
42 | use B (); |
54 | use B (); |
43 | #use Digest::MD5 (); |
|
|
44 | use DynaLoader (); |
55 | use DynaLoader (); |
|
|
56 | use Digest::MD5 (); |
|
|
57 | use Storable (); |
|
|
58 | use Fcntl (); |
45 | |
59 | |
46 | BEGIN { |
60 | BEGIN { |
47 | our $VERSION = '0.01'; |
61 | our $VERSION = '0.1'; |
48 | |
62 | |
49 | require XSLoader; |
63 | require XSLoader; |
50 | XSLoader::load __PACKAGE__, $VERSION; |
64 | XSLoader::load __PACKAGE__, $VERSION; |
51 | } |
65 | } |
52 | |
66 | |
|
|
67 | my $CACHEDIR = |
|
|
68 | $ENV{FASTER_CACHE} |
|
|
69 | || (exists $ENV{HOME} && "$ENV{HOME}/.perl-faster-cache") |
|
|
70 | || do { |
|
|
71 | require File::Temp; |
|
|
72 | File::Temp::tempdir (CLEANUP => 1) |
|
|
73 | }; |
|
|
74 | |
53 | my $COMPILE = "$Config{cc} -c -I$Config{archlibexp}/CORE $Config{optimize} $Config{ccflags} $Config{cccdlflags}"; |
75 | my $COMPILE = "$Config{cc} -c -I$Config{archlibexp}/CORE $Config{optimize} $Config{ccflags} $Config{cccdlflags}"; |
54 | my $LINK = "$Config{ld} $Config{ldflags} $Config{lddlflags} $Config{ccdlflags}"; |
76 | my $LINK = "$Config{ld} $Config{ldflags} $Config{lddlflags} $Config{ccdlflags}"; |
55 | my $LIBS = "$Config{libs}"; |
77 | my $LIBS = ""; |
56 | my $_o = $Config{_o}; |
78 | my $_o = $Config{_o}; |
57 | my $_so = ".so"; |
79 | my $_so = ".so"; |
58 | |
80 | |
59 | # we don't need no steenking PIC on x86 |
81 | # we don't need no steenking PIC on x86 |
60 | $COMPILE =~ s/-f(?:PIC|pic)//g |
82 | $COMPILE =~ s/-f(?:PIC|pic)//g |
61 | if $Config{archname} =~ /^(i[3456]86)-/; |
83 | if $Config{archname} =~ /^(i[3456]86)-/; |
62 | |
84 | |
63 | my $opt_assert = 0; |
85 | my $opt_assert = $ENV{FASTER_DEBUG} & 2; |
|
|
86 | my $verbose = $ENV{FASTER_VERBOSE}+0; |
|
|
87 | |
|
|
88 | warn "Faster: CACHEDIR is $CACHEDIR\n" if $verbose > 2; |
64 | |
89 | |
65 | our $source; |
90 | our $source; |
66 | |
91 | |
67 | our @ops; |
92 | our @ops; |
|
|
93 | our $insn; |
68 | our $op; |
94 | our $op; |
69 | our $op_name; |
95 | our $op_name; |
70 | our @op_loop; |
|
|
71 | our %op_regcomp; |
96 | our %op_regcomp; |
72 | |
97 | |
|
|
98 | # ops that cause immediate return to the interpreter |
73 | my %f_unsafe = map +($_ => undef), qw( |
99 | my %f_unsafe = map +($_ => undef), qw( |
74 | leavesub leavesublv return |
100 | leavesub leavesublv return |
75 | goto last redo next |
101 | goto last redo next |
76 | eval flip leaveeval entertry |
102 | eval flip leaveeval entertry |
77 | formline grepstart mapstart |
103 | formline grepstart mapstart |
78 | substcont entereval require |
104 | substcont entereval require |
79 | ); |
105 | ); |
80 | |
106 | |
81 | # pushmark extend=0 |
107 | # ops with known stack extend behaviour |
82 | # padsv extend=1 |
108 | # the values given are maximum values |
83 | # padav extend=1 |
109 | my %extend = ( |
84 | # padhv extend=1 |
110 | pushmark => 0, |
85 | # padany extend=1 |
111 | nextstate => 0, # might reduce the stack |
86 | # const extend=1 |
112 | unstack => 0, |
|
|
113 | enter => 0, |
87 | |
114 | |
|
|
115 | stringify => 0, |
|
|
116 | not => 0, |
|
|
117 | and => 0, |
|
|
118 | or => 0, |
|
|
119 | gvsv => 0, |
|
|
120 | rv2gv => 0, |
|
|
121 | preinc => 0, |
|
|
122 | predec => 0, |
|
|
123 | postinc => 0, |
|
|
124 | postdec => 0, |
|
|
125 | aelem => 0, |
|
|
126 | helem => 0, |
|
|
127 | qr => 1, #??? |
|
|
128 | pushre => 1, |
|
|
129 | gv => 1, |
|
|
130 | aelemfast => 1, |
|
|
131 | aelem => 0, |
|
|
132 | padsv => 1, |
|
|
133 | const => 1, |
|
|
134 | pop => 1, |
|
|
135 | shift => 1, |
|
|
136 | eq => -1, |
|
|
137 | ne => -1, |
|
|
138 | gt => -1, |
|
|
139 | lt => -1, |
|
|
140 | ge => -1, |
|
|
141 | lt => -1, |
|
|
142 | cond_expr => -1, |
|
|
143 | add => -1, |
|
|
144 | subtract => -1, |
|
|
145 | multiply => -1, |
|
|
146 | divide => -1, |
|
|
147 | aassign => 0, |
|
|
148 | sassign => -2, |
|
|
149 | method => 0, |
|
|
150 | method_named => 1, |
|
|
151 | ); |
|
|
152 | |
|
|
153 | # ops that do not need an ASYNC_CHECK |
88 | my %f_noasync = map +($_ => undef), qw( |
154 | my %f_noasync = map +($_ => undef), qw( |
89 | mapstart grepstart match entereval |
155 | mapstart grepstart match entereval |
90 | enteriter entersub leaveloop |
156 | enteriter entersub leaveloop |
91 | |
157 | |
92 | pushmark nextstate |
158 | pushmark nextstate caller |
93 | |
159 | |
94 | const stub unstack |
160 | const stub unstack |
95 | last next redo seq |
161 | last next redo goto seq |
96 | padsv padav padhv padany |
162 | padsv padav padhv padany |
97 | aassign sassign orassign |
163 | aassign sassign orassign |
98 | rv2av rv2cv rv2gv rv2hv refgen |
164 | rv2av rv2cv rv2gv rv2hv refgen |
99 | gv gvsv |
165 | gv gvsv |
100 | add subtract multiply divide |
166 | add subtract multiply divide |
101 | complement cond_expr and or not |
167 | complement cond_expr and or not |
|
|
168 | bit_and bit_or bit_xor |
102 | defined |
169 | defined |
103 | method_named |
170 | method method_named bless |
104 | preinc postinc predec postdec |
171 | preinc postinc predec postdec |
105 | aelem aelemfast helem delete exists |
172 | aelem aelemfast helem delete exists |
106 | pushre subst list join split concat |
173 | pushre subst list lslice join split concat |
107 | length substr stringify ord |
174 | length substr stringify ord |
108 | push pop shift unshift |
175 | push pop shift unshift |
109 | eq ne gt lt ge le |
176 | eq ne gt lt ge le |
110 | regcomp regcreset regcmaybe |
177 | regcomp regcreset regcmaybe |
111 | ); |
178 | ); |
112 | |
179 | |
113 | my %callop = ( |
180 | my %callop = ( |
114 | entersub => "(PL_ppaddr [OP_ENTERSUB]) (aTHX)", |
181 | entersub => "(PL_op->op_ppaddr) (aTHX)", |
115 | mapstart => "Perl_pp_grepstart (aTHX)", |
182 | mapstart => "Perl_pp_grepstart (aTHX)", |
116 | ); |
183 | ); |
117 | |
184 | |
118 | sub callop { |
185 | sub callop { |
119 | $callop{$op_name} || "Perl_pp_$op_name (aTHX)" |
186 | $callop{$op_name} || "Perl_pp_$op_name (aTHX)" |
… | |
… | |
125 | } |
192 | } |
126 | |
193 | |
127 | sub out_callop { |
194 | sub out_callop { |
128 | assert "nextop == (OP *)$$op"; |
195 | assert "nextop == (OP *)$$op"; |
129 | $source .= " PL_op = nextop; nextop = " . (callop $op) . ";\n"; |
196 | $source .= " PL_op = nextop; nextop = " . (callop $op) . ";\n"; |
|
|
197 | } |
|
|
198 | |
|
|
199 | sub out_jump { |
|
|
200 | assert "nextop == (OP *)${$_[0]}L"; |
|
|
201 | $source .= " goto op_${$_[0]};\n"; |
130 | } |
202 | } |
131 | |
203 | |
132 | sub out_cond_jump { |
204 | sub out_cond_jump { |
133 | $source .= " if (nextop == (OP *)${$_[0]}L) goto op_${$_[0]};\n"; |
205 | $source .= " if (nextop == (OP *)${$_[0]}L) goto op_${$_[0]};\n"; |
134 | } |
206 | } |
… | |
… | |
167 | |
239 | |
168 | out_next; |
240 | out_next; |
169 | } |
241 | } |
170 | |
242 | |
171 | sub op_pushmark { |
243 | sub op_pushmark { |
172 | $source .= " PUSHMARK (PL_stack_sp);\n"; |
244 | $source .= " faster_PUSHMARK (PL_stack_sp);\n"; |
173 | |
245 | |
174 | out_next; |
246 | out_next; |
175 | } |
247 | } |
176 | |
248 | |
177 | if ($Config{useithreads} ne "define") { |
249 | if ($Config{useithreads} ne "define") { |
178 | # disable optimisations on ithreads |
250 | # disable optimisations on ithreads |
179 | |
251 | |
180 | *op_const = sub { |
252 | *op_const = sub { |
181 | $source .= " { dSP; XPUSHs ((SV *)${$op->sv}L); PUTBACK; }\n"; |
253 | $source .= " { dSP; PUSHs ((SV *)${$op->sv}L); PUTBACK; }\n"; |
|
|
254 | |
|
|
255 | $ops[0]{follows_const}++ if @ops;#d# |
182 | |
256 | |
183 | out_next; |
257 | out_next; |
184 | }; |
258 | }; |
185 | |
259 | |
186 | *op_gv = \&op_const; |
260 | *op_gv = \&op_const; |
… | |
… | |
206 | if (!($op->flags & B::OPf_MOD)) { |
280 | if (!($op->flags & B::OPf_MOD)) { |
207 | $source .= " if (SvGMAGICAL (sv)) sv = sv_mortalcopy (sv);\n"; |
281 | $source .= " if (SvGMAGICAL (sv)) sv = sv_mortalcopy (sv);\n"; |
208 | } |
282 | } |
209 | |
283 | |
210 | $source .= " dSP;\n"; |
284 | $source .= " dSP;\n"; |
211 | $source .= " XPUSHs (sv);\n"; |
285 | $source .= " PUSHs (sv);\n"; |
212 | $source .= " PUTBACK;\n"; |
286 | $source .= " PUTBACK;\n"; |
213 | $source .= " }\n"; |
287 | $source .= " }\n"; |
214 | |
288 | |
215 | out_next; |
289 | out_next; |
216 | }; |
290 | }; |
217 | |
291 | |
218 | *op_gvsv = sub { |
292 | *op_gvsv = sub { |
219 | $source .= " {\n"; |
293 | $source .= " {\n"; |
220 | $source .= " dSP;\n"; |
294 | $source .= " dSP;\n"; |
221 | $source .= " EXTEND (SP, 1);\n"; |
|
|
222 | |
295 | |
223 | if ($op->private & B::OPpLVAL_INTRO) { |
296 | if ($op->private & B::OPpLVAL_INTRO) { |
224 | $source .= " PUSHs (save_scalar ((GV *)${$op->sv}L));\n"; |
297 | $source .= " PUSHs (save_scalar ((GV *)${$op->sv}L));\n"; |
225 | } else { |
298 | } else { |
226 | $source .= " PUSHs (GvSV ((GV *)${$op->sv}L));\n"; |
299 | $source .= " PUSHs (GvSV ((GV *)${$op->sv}L));\n"; |
… | |
… | |
286 | out_next; |
359 | out_next; |
287 | } |
360 | } |
288 | |
361 | |
289 | sub op_padsv { |
362 | sub op_padsv { |
290 | my $flags = $op->flags; |
363 | my $flags = $op->flags; |
291 | my $target = $op->targ; |
364 | my $padofs = "(PADOFFSET)" . $op->targ; |
292 | |
365 | |
293 | $source .= <<EOF; |
366 | $source .= <<EOF; |
294 | { |
367 | { |
295 | dSP; |
368 | dSP; |
296 | XPUSHs (PAD_SV ((PADOFFSET)$target)); |
369 | SV *sv = PAD_SVl ($padofs); |
|
|
370 | EOF |
|
|
371 | |
|
|
372 | if (($flags & B::OPf_MOD) && ($op->private & B::OPpLVAL_INTRO)) { |
|
|
373 | $source .= " SAVECLEARSV (PAD_SVl ($padofs));\n"; |
|
|
374 | $ops[0]{follows_padsv_lval_intro}++ if @ops;#d# |
|
|
375 | } |
|
|
376 | |
|
|
377 | $source .= <<EOF; |
|
|
378 | PUSHs (sv); |
297 | PUTBACK; |
379 | PUTBACK; |
298 | EOF |
380 | EOF |
299 | if ($op->flags & B::OPf_MOD) { |
381 | |
300 | if ($op->private & B::OPpLVAL_INTRO) { |
382 | if (($flags & B::OPf_MOD) && ($op->private & B::OPpDEREF)) { |
301 | $source .= " SAVECLEARSV (PAD_SVl ((PADOFFSET)$target));\n"; |
383 | $source .= " if (!SvROK (sv)) vivify_ref (sv, " . $op->private . " & OPpDEREF);\n"; |
302 | } elsif ($op->private & B::OPpDEREF) { |
|
|
303 | my $deref = $op->private & B::OPpDEREF; |
|
|
304 | $source .= " Perl_vivify_ref (PAD_SVl ((PADOFFSET)$target), $deref);\n"; |
|
|
305 | } |
|
|
306 | } |
384 | } |
|
|
385 | $source .= " }\n"; |
|
|
386 | |
|
|
387 | out_next; |
|
|
388 | } |
|
|
389 | |
|
|
390 | sub op_sassign { |
|
|
391 | $source .= <<EOF; |
|
|
392 | { |
|
|
393 | dSP; |
|
|
394 | dPOPTOPssrl; |
|
|
395 | EOF |
|
|
396 | $source .= " SV *temp = left; left = right; right = temp;\n" |
|
|
397 | if $op->private & B::OPpASSIGN_BACKWARDS; |
|
|
398 | |
|
|
399 | if ($insn->{follows_padsv_lval_intro} && !($op->private & B::OPpASSIGN_BACKWARDS)) { |
|
|
400 | # simple assignment - the target exists, but is basically undef |
|
|
401 | $source .= " SvSetSV (right, left);\n"; |
|
|
402 | } else { |
|
|
403 | $source .= " SvSetMagicSV (right, left);\n"; |
|
|
404 | } |
|
|
405 | |
307 | $source .= <<EOF; |
406 | $source .= <<EOF; |
|
|
407 | SETs (right); |
|
|
408 | PUTBACK; |
308 | } |
409 | } |
309 | EOF |
410 | EOF |
310 | |
411 | |
311 | out_next; |
412 | out_next; |
312 | } |
413 | } |
313 | |
414 | |
314 | # pattern const+ (or general push1) |
415 | # pattern const+ (or general push1) |
315 | # pattern pushmark return(?) |
|
|
316 | # pattern pushmark gv rv2av pushmark padsv+o.ä. aassign |
416 | # pattern pushmark gv rv2av pushmark padsv+o.ä. aassign |
317 | |
417 | |
318 | # pattern const method_named |
|
|
319 | sub op_method_named { |
418 | sub op_method_named { |
|
|
419 | if ($insn->{follows_const}) { |
320 | $source .= <<EOF; |
420 | $source .= <<EOF; |
|
|
421 | { |
|
|
422 | dSP; |
|
|
423 | static SV *last_cv; |
|
|
424 | static U32 last_sub_generation; |
|
|
425 | |
|
|
426 | /* simple "polymorphic" inline cache */ |
|
|
427 | if (PL_sub_generation == last_sub_generation) |
|
|
428 | { |
|
|
429 | PUSHs (last_cv); |
|
|
430 | PUTBACK; |
|
|
431 | } |
|
|
432 | else |
|
|
433 | { |
|
|
434 | PL_op = nextop; nextop = Perl_pp_method_named (aTHX); |
|
|
435 | |
|
|
436 | SPAGAIN; |
|
|
437 | last_sub_generation = PL_sub_generation; |
|
|
438 | last_cv = TOPs; |
|
|
439 | } |
|
|
440 | } |
|
|
441 | EOF |
|
|
442 | } else { |
|
|
443 | $source .= <<EOF; |
321 | { |
444 | { |
322 | static HV *last_stash; |
445 | static HV *last_stash; |
323 | static SV *last_cv; |
446 | static SV *last_cv; |
324 | static U32 last_sub_generation; |
447 | static U32 last_sub_generation; |
325 | |
448 | |
… | |
… | |
332 | |
455 | |
333 | /* simple "polymorphic" inline cache */ |
456 | /* simple "polymorphic" inline cache */ |
334 | if (stash == last_stash |
457 | if (stash == last_stash |
335 | && PL_sub_generation == last_sub_generation) |
458 | && PL_sub_generation == last_sub_generation) |
336 | { |
459 | { |
337 | XPUSHs (last_cv); |
460 | PUSHs (last_cv); |
338 | PUTBACK; |
461 | PUTBACK; |
339 | } |
462 | } |
340 | else |
463 | else |
341 | { |
464 | { |
342 | PL_op = nextop; nextop = Perl_pp_method_named (aTHX); |
465 | PL_op = nextop; nextop = Perl_pp_method_named (aTHX); |
… | |
… | |
352 | /* error case usually */ |
475 | /* error case usually */ |
353 | PL_op = nextop; nextop = Perl_pp_method_named (aTHX); |
476 | PL_op = nextop; nextop = Perl_pp_method_named (aTHX); |
354 | } |
477 | } |
355 | } |
478 | } |
356 | EOF |
479 | EOF |
|
|
480 | } |
357 | |
481 | |
358 | out_next; |
482 | out_next; |
359 | } |
483 | } |
360 | |
484 | |
361 | sub op_grepstart { |
485 | sub op_grepstart { |
… | |
… | |
375 | } |
499 | } |
376 | |
500 | |
377 | sub out_break_op { |
501 | sub out_break_op { |
378 | my ($idx) = @_; |
502 | my ($idx) = @_; |
379 | |
503 | |
|
|
504 | if ($op->flags & B::OPf_SPECIAL && $insn->{loop}) { |
|
|
505 | # common case: no label, innermost loop only |
|
|
506 | my $next = $insn->{loop}{loop_targ}[$idx]; |
380 | out_callop; |
507 | out_callop; |
381 | |
508 | out_jump $next; |
382 | out_cond_jump $_->[$idx] |
509 | } elsif (my $loop = $insn->{loop}) { |
383 | for reverse @op_loop; |
510 | # less common case: maybe break to some outer loop |
384 | |
|
|
385 | $source .= " return nextop;\n"; |
511 | $source .= " return nextop;\n"; |
|
|
512 | # todo: walk stack up |
|
|
513 | } else { |
|
|
514 | # fuck yourself for writing such hacks |
|
|
515 | $source .= " return nextop;\n"; |
|
|
516 | } |
386 | } |
517 | } |
387 | |
518 | |
388 | sub xop_next { |
519 | sub op_next { |
389 | out_break_op 0; |
520 | out_break_op 0; |
390 | } |
521 | } |
391 | |
522 | |
392 | sub op_last { |
523 | sub op_last { |
393 | out_break_op 1; |
524 | out_break_op 1; |
394 | } |
525 | } |
395 | |
526 | |
|
|
527 | # TODO: does not seem to work |
396 | sub xop_redo { |
528 | #sub op_redo { |
397 | out_break_op 2; |
529 | # out_break_op 2; |
398 | } |
530 | #} |
399 | |
531 | |
400 | sub cv2c { |
532 | sub cv2c { |
401 | my ($cv) = @_; |
533 | my ($cv) = @_; |
402 | |
534 | |
403 | local @ops; |
535 | local @ops; |
404 | local @op_loop; |
|
|
405 | local %op_regcomp; |
536 | local %op_regcomp; |
406 | |
537 | |
407 | my %opsseen; |
538 | my $curloop; |
408 | my @todo = $cv->START; |
539 | my @todo = $cv->START; |
|
|
540 | my %op_target; |
|
|
541 | my $numpushmark; |
|
|
542 | my $scope; |
409 | |
543 | |
|
|
544 | my %op_seen; |
410 | while (my $op = shift @todo) { |
545 | while (my $op = shift @todo) { |
|
|
546 | my $next; |
411 | for (; $$op; $op = $op->next) { |
547 | for (; $$op; $op = $next) { |
412 | last if $opsseen{$$op}++; |
548 | last if $op_seen{$$op}++; |
413 | push @ops, $op; |
549 | |
|
|
550 | $next = $op->next; |
414 | |
551 | |
415 | my $name = $op->name; |
552 | my $name = $op->name; |
416 | my $class = B::class $op; |
553 | my $class = B::class $op; |
417 | |
554 | |
|
|
555 | my $insn = { op => $op }; |
|
|
556 | |
|
|
557 | # end of loop reached? |
|
|
558 | $curloop = $curloop->{loop} if $curloop && $$op == ${$curloop->{loop_targ}[1]}; |
|
|
559 | |
|
|
560 | # remember enclosing loop |
|
|
561 | $insn->{loop} = $curloop if $curloop; |
|
|
562 | |
|
|
563 | push @ops, $insn; |
|
|
564 | |
|
|
565 | if (exists $extend{$name}) { |
|
|
566 | my $extend = $extend{$name}; |
|
|
567 | $extend = $extend->($op) if ref $extend; |
|
|
568 | $insn->{extend} = $extend if defined $extend; |
|
|
569 | } |
|
|
570 | |
|
|
571 | # TODO: mark scopes similar to loops, make them comparable |
|
|
572 | # static cxstack(?) |
418 | if ($class eq "LOGOP") { |
573 | if ($class eq "LOGOP") { |
419 | unshift @todo, $op->other; # unshift vs. push saves jumps |
574 | push @todo, $op->other; |
|
|
575 | $op_target{${$op->other}}++; |
420 | |
576 | |
421 | # regcomp/o patches ops at runtime, lets expect that |
577 | # regcomp/o patches ops at runtime, lets expect that |
|
|
578 | if ($name eq "regcomp" && $op->other->pmflags & B::PMf_KEEP) { |
|
|
579 | $op_target{${$op->first}}++; |
422 | $op_regcomp{${$op->first}} = $op->next |
580 | $op_regcomp{${$op->first}} = $op->next; |
423 | if $name eq "regcomp" && $op->other->pmflags & B::PMf_KEEP; |
581 | } |
424 | |
582 | |
425 | } elsif ($class eq "PMOP") { |
583 | } elsif ($class eq "PMOP") { |
|
|
584 | if (${$op->pmreplstart}) { |
426 | unshift @todo, $op->pmreplstart if ${$op->pmreplstart}; |
585 | unshift @todo, $op->pmreplstart; |
|
|
586 | $op_target{${$op->pmreplstart}}++; |
|
|
587 | } |
427 | |
588 | |
428 | } elsif ($class eq "LOOP") { |
589 | } elsif ($class eq "LOOP") { |
429 | push @op_loop, [$op->nextop, $op->lastop->next, $op->redoop->next]; |
|
|
430 | push @todo, $op->nextop, $op->lastop->next, $op->redoop->next; |
590 | my @targ = ($op->nextop, $op->lastop->next, $op->redoop); |
|
|
591 | |
|
|
592 | unshift @todo, $next, $op->redoop, $op->nextop, $op->lastop; |
|
|
593 | $next = $op->redoop; |
|
|
594 | |
|
|
595 | $op_target{$$_}++ for @targ; |
|
|
596 | |
|
|
597 | $insn->{loop_targ} = \@targ; |
|
|
598 | $curloop = $insn; |
|
|
599 | |
|
|
600 | } elsif ($class eq "COP") { |
|
|
601 | if (defined $op->label) { |
|
|
602 | $insn->{bblock}++; |
|
|
603 | $curloop->{contains_label}{$op->label}++ if $curloop; #TODO: should be within loop |
|
|
604 | } |
|
|
605 | |
|
|
606 | } else { |
|
|
607 | if ($name eq "pushmark") { |
|
|
608 | $numpushmark++; |
|
|
609 | } |
431 | } |
610 | } |
432 | } |
611 | } |
433 | } |
612 | } |
|
|
613 | |
|
|
614 | $_->{bblock}++ for grep $op_target{${$_->{op}}}, @ops; |
434 | |
615 | |
435 | local $source = <<EOF; |
616 | local $source = <<EOF; |
436 | OP *%%%FUNC%%% (pTHX) |
617 | OP *%%%FUNC%%% (pTHX) |
437 | { |
618 | { |
438 | register OP *nextop = (OP *)${$ops[0]}L; |
619 | register OP *nextop = (OP *)${$ops[0]->{op}}L; |
439 | EOF |
620 | EOF |
|
|
621 | |
|
|
622 | $source .= " faster_PUSHMARK_PREALLOC ($numpushmark);\n" |
|
|
623 | if $numpushmark; |
440 | |
624 | |
441 | while (@ops) { |
625 | while (@ops) { |
442 | $op = shift @ops; |
626 | $insn = shift @ops; |
|
|
627 | |
|
|
628 | $op = $insn->{op}; |
443 | $op_name = $op->name; |
629 | $op_name = $op->name; |
444 | |
630 | |
|
|
631 | my $class = B::class $op; |
|
|
632 | |
|
|
633 | $source .= "\n/* start basic block */\n" if exists $insn->{bblock};#d# |
445 | $source .= "op_$$op: /* $op_name */\n"; |
634 | $source .= "op_$$op: /* $op_name */\n"; |
446 | #$source .= "fprintf (stderr, \"$$op in op $op_name\\n\");\n";#d# |
635 | #$source .= "fprintf (stderr, \"$$op in op $op_name\\n\");\n";#d# |
447 | #$source .= "{ dSP; sv_dump (TOPs); }\n";#d# |
636 | #$source .= "{ dSP; sv_dump (TOPs); }\n";#d# |
448 | |
637 | |
449 | $source .= " PERL_ASYNC_CHECK ();\n" |
638 | $source .= " PERL_ASYNC_CHECK ();\n" |
450 | unless exists $f_noasync{$op_name}; |
639 | unless exists $f_noasync{$op_name}; |
451 | |
640 | |
452 | if (my $can = __PACKAGE__->can ("op_$op_name")) { |
641 | if (my $can = __PACKAGE__->can ("op_$op_name")) { |
453 | # handcrafted replacement |
642 | # handcrafted replacement |
|
|
643 | |
|
|
644 | if ($insn->{extend} > 0) { |
|
|
645 | # coalesce EXTENDs |
|
|
646 | # TODO: properly take negative preceeding and following EXTENDs into account |
|
|
647 | for my $i (@ops) { |
|
|
648 | last if exists $i->{bblock}; |
|
|
649 | last unless exists $i->{extend}; |
|
|
650 | my $extend = delete $i->{extend}; |
|
|
651 | $insn->{extend} += $extend if $extend > 0; |
|
|
652 | } |
|
|
653 | |
|
|
654 | $source .= " { dSP; EXTEND (SP, $insn->{extend}); PUTBACK; }\n" |
|
|
655 | if $insn->{extend} > 0; |
|
|
656 | } |
|
|
657 | |
454 | $can->($op); |
658 | $can->($op); |
455 | |
659 | |
456 | } elsif (exists $f_unsafe{$op_name}) { |
660 | } elsif (exists $f_unsafe{$op_name}) { |
457 | # unsafe, return to interpreter |
661 | # unsafe, return to interpreter |
458 | assert "nextop == (OP *)$$op"; |
662 | assert "nextop == (OP *)$$op"; |
459 | $source .= " return nextop;\n"; |
663 | $source .= " return nextop;\n"; |
460 | |
664 | |
461 | } elsif ("LOGOP" eq B::class $op) { |
665 | } elsif ("LOGOP" eq $class) { |
462 | # logical operation with optionaö branch |
666 | # logical operation with optional branch |
463 | out_callop; |
667 | out_callop; |
464 | out_cond_jump $op->other; |
668 | out_cond_jump $op->other; |
465 | out_jump_next; |
669 | out_jump_next; |
466 | |
670 | |
467 | } elsif ("PMOP" eq B::class $op) { |
671 | } elsif ("PMOP" eq $class) { |
468 | # regex-thingy |
672 | # regex-thingy |
469 | out_callop; |
673 | out_callop; |
470 | out_cond_jump $op->pmreplroot if ${$op->pmreplroot}; |
674 | out_cond_jump $op->pmreplroot if $op_name ne "pushre" && ${$op->pmreplroot}; |
471 | out_jump_next; |
675 | out_jump_next; |
472 | |
676 | |
473 | } else { |
677 | } else { |
474 | # normal operator, linear execution |
678 | # normal operator, linear execution |
475 | out_linear; |
679 | out_linear; |
… | |
… | |
487 | |
691 | |
488 | $source |
692 | $source |
489 | } |
693 | } |
490 | |
694 | |
491 | my $uid = "aaaaaaa0"; |
695 | my $uid = "aaaaaaa0"; |
|
|
696 | my %so; |
492 | |
697 | |
493 | sub source2ptr { |
698 | sub func2ptr { |
494 | my (@source) = @_; |
699 | my (@func) = @_; |
495 | |
700 | |
496 | my $stem = "/tmp/Faster-$$-" . $uid++; |
701 | #LOCK |
|
|
702 | mkdir $CACHEDIR, 0777; |
|
|
703 | sysopen my $meta_fh, "$CACHEDIR/meta", &Fcntl::O_RDWR | &Fcntl::O_CREAT, 0666 |
|
|
704 | or die "$$CACHEDIR/meta: $!"; |
|
|
705 | binmode $meta_fh, ":raw:perlio"; |
|
|
706 | fcntl_lock fileno $meta_fh |
|
|
707 | or die "$CACHEDIR/meta: $!"; |
497 | |
708 | |
|
|
709 | my $meta = eval { Storable::fd_retrieve $meta_fh } || { version => 1 }; |
|
|
710 | |
|
|
711 | for my $f (@func) { |
|
|
712 | $f->{func} = "F" . Digest::MD5::md5_hex ($f->{source}); |
|
|
713 | $f->{so} = $meta->{$f->{func}}; |
|
|
714 | } |
|
|
715 | |
|
|
716 | if (grep !$_->{so}, @func) { |
|
|
717 | my $stem; |
|
|
718 | |
|
|
719 | do { |
|
|
720 | $stem = "$CACHEDIR/$$-" . $uid++; |
|
|
721 | } while -e "$stem$_so"; |
|
|
722 | |
498 | open FILE, ">:raw", "$stem.c"; |
723 | open my $fh, ">:raw", "$stem.c"; |
499 | print FILE <<EOF; |
724 | print $fh <<EOF; |
500 | #define PERL_NO_GET_CONTEXT |
725 | #define PERL_NO_GET_CONTEXT |
|
|
726 | #define PERL_CORE |
501 | |
727 | |
502 | #include <assert.h> |
728 | #include <assert.h> |
503 | |
729 | |
504 | #include "EXTERN.h" |
730 | #include "EXTERN.h" |
505 | #include "perl.h" |
731 | #include "perl.h" |
506 | #include "XSUB.h" |
732 | #include "XSUB.h" |
507 | |
733 | |
|
|
734 | #if 1 |
|
|
735 | # define faster_PUSHMARK_PREALLOC(count) while (PL_markstack_ptr + (count) >= PL_markstack_max) markstack_grow () |
|
|
736 | # define faster_PUSHMARK(p) *++PL_markstack_ptr = (p) - PL_stack_base |
|
|
737 | #else |
|
|
738 | # define faster_PUSHMARK_PREALLOC(count) 1 |
|
|
739 | # define faster_PUSHMARK(p) PUSHMARK(p) |
|
|
740 | #endif |
|
|
741 | |
508 | #define RUNOPS_TILL(op) \\ |
742 | #define RUNOPS_TILL(op) \\ |
509 | while (nextop != (op)) \\ |
743 | while (nextop != (op)) \\ |
510 | { \\ |
744 | { \\ |
511 | PERL_ASYNC_CHECK (); \\ |
745 | PERL_ASYNC_CHECK (); \\ |
512 | PL_op = nextop; nextop = (PL_op->op_ppaddr)(aTHX); \\ |
746 | PL_op = nextop; nextop = (PL_op->op_ppaddr)(aTHX); \\ |
513 | } |
|
|
514 | |
|
|
515 | EOF |
|
|
516 | for (@source) { |
|
|
517 | my $func = $uid++; |
|
|
518 | $_ =~ s/%%%FUNC%%%/$func/g; |
|
|
519 | print FILE $_; |
|
|
520 | $_ = $func; |
|
|
521 | } |
747 | } |
522 | |
748 | |
523 | close FILE; |
749 | EOF |
|
|
750 | for my $f (grep !$_->{so}, @func) { |
|
|
751 | next if $f->{so} = $meta->{$f->{func}}; # some cv's alias others |
|
|
752 | |
|
|
753 | warn "compiling $f->{name} to $stem$_so:$f->{func}\n" if $verbose > 1; |
|
|
754 | my $source = $f->{source}; |
|
|
755 | $source =~ s/%%%FUNC%%%/$f->{func}/g; |
|
|
756 | print $fh $source; |
|
|
757 | $meta->{$f->{func}} = $f->{so} = $stem; |
|
|
758 | } |
|
|
759 | |
|
|
760 | close $fh; |
524 | system "$COMPILE -o $stem$_o $stem.c"; |
761 | system "$COMPILE -o $stem$_o $stem.c"; |
525 | #d#unlink "$stem.c"; |
762 | unlink "$stem.c" unless $ENV{FASTER_DEBUG} & 1; |
526 | system "$LINK -o $stem$_so $stem$_o $LIBS"; |
763 | system "$LINK -o $stem$_so $stem$_o $LIBS"; |
527 | unlink "$stem$_o"; |
764 | unlink "$stem$_o"; |
|
|
765 | } |
528 | |
766 | |
|
|
767 | for my $f (@func) { |
|
|
768 | my $stem = $f->{so}; |
|
|
769 | |
529 | my $so = DynaLoader::dl_load_file "$stem$_so" |
770 | my $so = ($so{$stem} ||= DynaLoader::dl_load_file "$stem$_so") |
530 | or die "$stem$_so: $!"; |
771 | or die "$stem$_so: $!"; |
531 | |
772 | |
532 | #unlink "$stem$_so"; |
773 | #unlink "$stem$_so"; |
533 | |
774 | |
534 | map +(DynaLoader::dl_find_symbol $so, $_), @source |
775 | $f->{ptr} = DynaLoader::dl_find_symbol $so, $f->{func} |
|
|
776 | or die "$f->{func} not found in $stem$_so: $!"; |
|
|
777 | } |
|
|
778 | |
|
|
779 | seek $meta_fh, 0, 0 or die "$CACHEDIR/meta: $!"; |
|
|
780 | Storable::nstore_fd $meta, $meta_fh; |
|
|
781 | truncate $meta_fh, tell $meta_fh; |
|
|
782 | |
|
|
783 | # UNLOCK (by closing $meta_fh) |
535 | } |
784 | } |
536 | |
785 | |
537 | my %ignore; |
786 | my %ignore; |
538 | |
787 | |
539 | sub entersub { |
788 | sub entersub { |
… | |
… | |
541 | |
790 | |
542 | my $pkg = $cv->STASH->NAME; |
791 | my $pkg = $cv->STASH->NAME; |
543 | |
792 | |
544 | return if $ignore{$pkg}; |
793 | return if $ignore{$pkg}; |
545 | |
794 | |
546 | warn "compiling ", $cv->STASH->NAME;#d# |
795 | warn "optimising ", $cv->STASH->NAME, "\n" |
|
|
796 | if $verbose; |
547 | |
797 | |
548 | eval { |
798 | eval { |
549 | my @cv; |
799 | my @func; |
550 | my @cv_source; |
800 | |
|
|
801 | push @func, { |
|
|
802 | cv => $cv, |
|
|
803 | name => "<>", |
|
|
804 | source => cv2c $cv, |
|
|
805 | }; |
551 | |
806 | |
552 | # always compile the whole stash |
807 | # always compile the whole stash |
553 | my %stash = $cv->STASH->ARRAY; |
808 | my %stash = $cv->STASH->ARRAY; |
554 | while (my ($k, $v) = each %stash) { |
809 | while (my ($k, $v) = each %stash) { |
555 | $v->isa (B::GV::) |
810 | $v->isa (B::GV::) |
… | |
… | |
558 | my $cv = $v->CV; |
813 | my $cv = $v->CV; |
559 | |
814 | |
560 | if ($cv->isa (B::CV::) |
815 | if ($cv->isa (B::CV::) |
561 | && ${$cv->START} |
816 | && ${$cv->START} |
562 | && $cv->START->name ne "null") { |
817 | && $cv->START->name ne "null") { |
|
|
818 | |
563 | push @cv, $cv; |
819 | push @func, { |
|
|
820 | cv => $cv, |
|
|
821 | name => $k, |
564 | push @cv_source, cv2c $cv; |
822 | source => cv2c $cv, |
|
|
823 | }; |
565 | } |
824 | } |
566 | } |
825 | } |
567 | |
826 | |
568 | my @ptr = source2ptr @cv_source; |
827 | func2ptr @func; |
569 | |
828 | |
570 | for (0 .. $#cv) { |
829 | for my $f (@func) { |
571 | patch_cv $cv[$_], $ptr[$_]; |
830 | patch_cv $f->{cv}, $f->{ptr}; |
572 | } |
831 | } |
573 | }; |
832 | }; |
574 | |
833 | |
575 | if ($@) { |
834 | if ($@) { |
576 | $ignore{$pkg}++; |
835 | $ignore{$pkg}++; |
… | |
… | |
582 | |
841 | |
583 | 1; |
842 | 1; |
584 | |
843 | |
585 | =back |
844 | =back |
586 | |
845 | |
|
|
846 | =head1 ENVIRONMENT VARIABLES |
|
|
847 | |
|
|
848 | The following environment variables influence the behaviour of Faster: |
|
|
849 | |
|
|
850 | =over 4 |
|
|
851 | |
|
|
852 | =item FASTER_VERBOSE |
|
|
853 | |
|
|
854 | Faster will output more informational messages when set to values higher |
|
|
855 | than C<0>. Currently, C<1> outputs which packages are being compiled, C<3> |
|
|
856 | outputs the cache directory and C<10> outputs information on which perl |
|
|
857 | function is compiled into which shared object. |
|
|
858 | |
|
|
859 | =item FASTER_DEBUG |
|
|
860 | |
|
|
861 | Add debugging code when set to values higher than C<0>. Currently, this |
|
|
862 | adds 1-3 C<assert>'s per perl op (FASTER_DEBUG > 1), to ensure that opcode |
|
|
863 | order and C execution order are compatible. |
|
|
864 | |
|
|
865 | =item FASTER_CACHE |
|
|
866 | |
|
|
867 | Set a persistent cache directory that caches compiled code fragments. The |
|
|
868 | default is C<$HOME/.perl-faster-cache> if C<HOME> is set and a temporary |
|
|
869 | directory otherwise. |
|
|
870 | |
|
|
871 | This directory will always grow in size, so you might need to erase it |
|
|
872 | from time to time. |
|
|
873 | |
|
|
874 | =back |
|
|
875 | |
587 | =head1 BUGS/LIMITATIONS |
876 | =head1 BUGS/LIMITATIONS |
588 | |
877 | |
589 | Perl will check much less often for asynchronous signals in |
878 | Perl will check much less often for asynchronous signals in |
590 | Faster-compiled code. It tries to check on every function call, loop |
879 | Faster-compiled code. It tries to check on every function call, loop |
591 | iteration and every I/O operator, though. |
880 | iteration and every I/O operator, though. |
… | |
… | |
603 | These constructs will force the use of the interpreter for the currently |
892 | These constructs will force the use of the interpreter for the currently |
604 | executed function as soon as they are being encountered during execution. |
893 | executed function as soon as they are being encountered during execution. |
605 | |
894 | |
606 | goto |
895 | goto |
607 | next, redo (but not well-behaved last's) |
896 | next, redo (but not well-behaved last's) |
|
|
897 | labels, if used |
608 | eval |
898 | eval |
609 | require |
899 | require |
610 | any use of formats |
900 | any use of formats |
611 | .., ... (flipflop operators) |
901 | .., ... (flipflop operators) |
612 | |
902 | |