NAME
perloptree - The Perl op tree
DESCRIPTION
Various material about the internal Perl compilation representation during parsing and optimization, before the actual execution begins. The "B" op tree.
The well-known perlguts.pod focuses more on the internal representation of the variables, but not so on the structure, the sequence and the optimization of the basic operations, the ops.
And we have perlhack.pod, which shows e.g. ways to hack into the op tree structure within the debugger. It focuses on getting people to start patching and hacking on the CORE, not understanding or writing compiler backends or optimizations, which the op tree mainly is used for. Btw., this here is merely prose around the source comments. But who reads source?
Brief Summary
The brief summary is very well described in the "perlguts#Compiled-code" in "Compiled-code" section of perlguts and at the top of op.c.
When Perl parses the source code (via Yacc perly.y), the so-called op tree, a tree of basic perl OP structs pointing to simple pp_opname functions, is generated bottom-up. Those pp_ functions - "PP Code" (for "Push / Pop Code") - have the same uniform API as the XS functions, all arguments and return values are transported on the stack. For example, an OP_CONST op points to the pp_const() function and to an SV containing the constant value. When pp_const() is executed, its job is to push that SV onto the stack.
OPs are created by the newFOO() functions, which are called from the parser (in perly.y) as the code is parsed. For example the Perl code $a + $b * $c would cause the equivalent of the following to be called (oversimplifying a bit):
newBINOP(OP_ADD, flags,
newSVREF($a),
newBINOP(OP_MULTIPLY, flags, newSVREF($b), newSVREF($c))
)See also "perlhack.pod#Op Trees"
The simplest type of an op structure is OP, a "BASEOP": this has no children. Unary operators, "UNOP"s, have one child, and this is pointed to by the op_first field. Binary operators ("BINOP"s) have not only an op_first field but also an op_last field. The most complex type of op is a "LISTOP", which has any number of children. In this case, the first child is pointed to by op_first and the last child by op_last. The children in between can be found by iteratively following the op_sibling pointer from the first child to the last.
There are also two other op types: a "PMOP" holds a regular expression, and has no children, and a "LOOP" may or may not have children. If the op_sibling field is non-zero, it behaves like a LISTOP. To complicate matters, if an UNOP is actually a null op after optimization (see "Compile pass 2: context propagation" below) it will still have children in accordance with its former type.
The beautiful thing about the op tree representation is that it is a strict 1:1 mapping to the actual source code, which is proven by the B::Deparse module, which generates readable source for the current op tree. Well, almost.
The Compiler
Perl's compiler is essentially a 3-pass compiler with interleaved phases:
1. A bottom-up pass
2. A top-down pass
3. An execution-order passCompile pass 1: check routines and constant folding
The bottom-up pass is represented by all the "newOP" routines and the ck_ routines. The bottom-upness is actually driven by yacc. So at the point that a ck_ routine fires, we have no idea what the context is, either upward in the syntax tree, or either forward or backward in the execution order. The bottom-up parser builds that part of the execution order it knows about, but if you follow the "next" links around, you'll find it's actually a closed loop through the top level node.
So when creating the ops in the first step, still bottom-up, for each op a check function (ck_ ()) is called, which which theroretically may destructively modify the whole tree, but because it knows almost nothing, it mostly just nullifies the current op. Or it might set the "op_next" pointer. See "Check Functions" for more.
Also, the subsequent constant folding routine fold_constants() may nullify certain ops, which are skipped during final execution. See "Constant Folding" for more.
Compile pass 2: context propagation
The context determines the type of the return value. When a context for a part of compile tree is known, it is propagated down through the tree. At this time the context can have 5 values (instead of 2 for runtime context): void, boolean, scalar, list, and lvalue. In contrast with the pass 1 this pass is processed from top to bottom: a node's context determines the context for its children.
Whenever the bottom-up parser gets to a node that supplies context to its components, it invokes that portion of the top-down pass that applies to that part of the subtree (and marks the top node as processed, so if a node further up supplies context, it doesn't have to take the plunge again). As a particular subcase of this, as the new node is built, it takes all the closed execution loops of its subcomponents and links them into a new closed loop for the higher level node. But it's still not the real execution order.
Todo: Sample where this context flag is stored
Additional context-dependent optimizations are performed at this time. Since at this moment the compile tree contains back-references (via "thread" pointers), nodes cannot be free()d now. To allow optimized-away nodes at this stage, such nodes are null()ified instead of free()'ing (i.e. their type is changed to OP_NULL).
Compile pass 3: peephole optimization
The actual execution order is not known till we get a grammar reduction to a top-level unit like a subroutine or file that will be called by "name" rather than via a "next" pointer. At that point, we can call into peep() to do that code's portion of the 3rd pass. It has to be recursive, but it's recursive on basic blocks, not on tree nodes.
So finally, when the full parse tree is generated, the "peephole optimizer" peep() is running. This pass is neither top-down or bottom-up, but in the execution order (with additional complications for conditionals).
This examines each op in the tree and attempts to determine "local" optimizations by "thinking ahead" one or two ops and seeing if multiple operations can be combined into one (by nullifying and re-ordering the next pointers).
It also checks for lexical issues such as the effect of use strict on bareword constants. Note that since the last walk the early sibling pointers for recursive (bottom-up) meta-inspection are useless, the final exec order is guaranteed by the next and flags fields.
basic vs exec order
The highly recursive Yacc parser generates the initial op tree in basic order. To save memory and run-time the final execution order of the ops in sequential order is not copied around, just the next pointers are rehooked in Perl_linklist() to the so-called exec order. So the exec walk through the linked-list of ops is not too cache-friendly.
In detail Perl_linklist() traverses the op tree, and sets op-next pointers to give the execution order for that op tree. op-sibling pointers are rarely unneeded after that.
Walkers can run in "basic" or "exec" order. "basic" is useful for the memory layout, it contains the history, "exec" is more useful to understand the logic and program flow. The "B::Bytecode" section has an extensive example about the order.
OP Structure and Inheritance
The basic struct op looks basically like
C<{ OP* op_next, OP* op_sibling, OP* op_ppaddr, ..., int op_flags, int op_private } OP;> See "BASEOP" below.
Each op is defined in size, arguments, return values, class and more in the opcode.pl table. (See "OP Class Declarations in opcode.pl" below.)
The class of an OP determines its size and the number of children. But the number and type of arguments is not so easy to declare as in C. opcode.pl tries to declare some XS-prototype like arguments, but in lisp we would say most ops are "special" functions, context-dependent, hard to parse.
B.pm http://search.cpan.org/perldoc?B contains these classes and inheritance:
@B::OP::ISA = 'B::OBJECT';
@B::UNOP::ISA = 'B::OP';
@B::BINOP::ISA = 'B::UNOP';
@B::LOGOP::ISA = 'B::UNOP';
@B::LISTOP::ISA = 'B::BINOP';
@B::SVOP::ISA = 'B::OP';
@B::PADOP::ISA = 'B::OP';
@B::PVOP::ISA = 'B::OP';
@B::LOOP::ISA = 'B::LISTOP';
@B::PMOP::ISA = 'B::LISTOP';
@B::COP::ISA = 'B::OP';
@B::SPECIAL::ISA = 'B::OBJECT';
@B::optype = qw(OP UNOP BINOP LOGOP LISTOP PMOP SVOP PADOP PVOP LOOP COP);TODO: ascii graph from perlguts
op.h http://search.cpan.org/src/RGARCIA/perl-5.10.0/op.h contains all the gory details. Let's check it out.
OP Class Declarations in opcode.pl
The full list of op declarations is defined as DATA in opcode.pl. It defines the class, the name, some flags, and the argument types, the so-called "operands". make regen (via regen.pl) recreates out of this DATA table the files opcode.h, opnames.h, pp_proto.h and pp.sym.
The class signifiers in opcode.pl are:
baseop - 0 unop - 1 binop - 2
logop - | listop - @ pmop - /
padop/svop - $ padop - # (unused) loop - {
baseop/unop - % loopexop - } filestatop - -
pvop/svop - " cop - ;Other options within opcode.pl are:
needs stack mark - m
needs constant folding - f
produces a scalar - s
produces an integer - i
needs a target - t
target can be in a pad - T
has a corresponding integer version - I
has side effects - d
uses $_ if no argument given - uValues for the operands are:
scalar - S list - L array - A
hash - H sub (CV) - C file - F
socket - Fs filetest - F- reference - R
"?" denotes an optional operand.BASEOP
All op classes have a single character signifier for easier definition in opcode.pl. The BASEOP class signifier is 0, for no children.
Below are the BASEOP fields, which reflect the object B::OP, since Perl 5.10. These are shared for all op classes. The parts after op_type and before op_flags changed during history.
- op_next
-
Pointer to next op to execute after this one.
Top level pre-grafted op points to first op, but this is replaced when op is grafted in, when this op will point to the real next op, and the new parent takes over role of remembering the starting op. Now, who wrote this prose? Anyway, that is why it is called guts.
- op_sibling
-
Pointer to connect the children's list.
The first child is "op_first", the last is "op_last", and the children in between are interconnected by op_sibling. This is at run-time only used for "LISTOP"s.
So why is it in the BASEOP struct carried around for every op?
Because of the complicated Yacc parsing and later optimization order as explained in "Compile pass 1: check routines and constant folding" the "op_next" pointers are not enough, so op_sibling's are required. The final and fast execution order by just following the op_next chain is expensive to calculate.
See http://www.xray.mpe.mpg.de/mailing-lists/perl5-porters/2006-09/msg00082.html for a 20% space-reduction patch to get rid of it at run-time.
- op_ppaddr
-
Pointer to current ppcode's function. The so called "opcode".
- op_madprop
-
Pointer to the MADPROP struct. Only with -DMAD, and since 5.10. See "MAD" (Misc Attribute Decoration) below.
- op_targ
-
PADOFFSET to "unnamed" op targets/GVs/constants, wasting no SV. Has for some op's also a different meaning.
- op_type
-
The type of the operation.
Since 5.10 we have the next five fields added, which replace
U16 op_seq. - op_opt
-
"optimized"
Whether or not the op has been optimised by the peephole optimiser.
See the comments in
S_clear_yystack()in perly.c for more details on the following three flags. They are just for freeing temporary ops on the stack. But we might have statically allocated op in the data segment, esp. with the perl compiler's B::C module. Then we are not allowed to free those static ops. For a short time, from 5.9.0 until 5.9.4, until the B::C module was removed from CORE, we had another field here for this reason: op_static. On 1 it didn't free the static op. Before 5.9.0 the "op_seq" field was used with the magic value -1 to indicate a static op, not to be freed. Note: Trying to free a static struct is considered harmful. - op_latefree
-
Tell
op_free()to clear this op (and free any kids) but not yet deallocate the struct. This means that the op may be safelyop_free()d multiple times.On static ops you just set this to 1 and after the first
op_free()theop_latefreedis automatically set and furtherop_free()called are just ignored. - op_latefreed
-
If 1, an
op_latefreeop has beenop_free()d. - op_attached
-
This op (sub)tree has been attached to the CV
PL_compcvso it doesn't need to be free'd. - op_spare
-
Three spare bits in this bitfield above. At least they survived 5.10.
Those last two fields have been in all perls:
- op_flags
-
Flags common to all operations. See
OPf_*in op.h, or more verbose in B::Flags or dump.c - op_private
-
Flags peculiar to a particular operation (BUT, by default, set to the number of children until the operation is privatized by a check routine, which may or may not check number of children).
This flag is normally used to hold op specific context hints, such as
HINT_INTEGER. This flag is directly attached to each relevant op in the subtree of the context. Note that there's no general context or class pointer for each op, a typical functional language usually holds this in the ops arguments. So we are limited to max 32 lexical pragma hints or less. See "Lexical Pragmas".
The exact op.h "BASEOP" history for the parts after op_type and before op_flags is:
<=5.8: U16 op_seq;
5.9.4: unsigned op_opt:1; unsigned op_static:1; unsigned op_spare:5;
>=5.10: unsigned op_opt:1; unsigned op_latefree:1; unsigned op_latefreed:1;
unsigned op_attached:1; unsigned op_spare:3;The "BASEOP" class signifier is 0, for no children. The full list of all BASEOP's is:
$ perl -F"/\cI+/" -ane 'print if $F[3] =~ /0$/' opcode.pl
null null operation ck_null 0
stub stub ck_null 0
pushmark pushmark ck_null s0
wantarray wantarray ck_null is0
padsv private variable ck_null ds0
padav private array ck_null d0
padhv private hash ck_null d0
padany private value ck_null d0
sassign scalar assignment ck_sassign s0
unstack iteration finalizer ck_null s0
enter block entry ck_null 0
iter foreach loop iterator ck_null 0
break break ck_null 0
continue continue ck_null 0
fork fork ck_null ist0
wait wait ck_null isT0
getppid getppid ck_null isT0
time time ck_null isT0
tms times ck_null 0
ghostent gethostent ck_null 0
gnetent getnetent ck_null 0
gprotoent getprotoent ck_null 0
gservent getservent ck_null 0
ehostent endhostent ck_null is0
enetent endnetent ck_null is0
eprotoent endprotoent ck_null is0
eservent endservent ck_null is0
gpwent getpwent ck_null 0
spwent setpwent ck_null is0
epwent endpwent ck_null is0
ggrent getgrent ck_null 0
sgrent setgrent ck_null is0
egrent endgrent ck_null is0
getlogin getlogin ck_null st0
custom unknown custom operator ck_null 0UNOP
The unary op class signifier is 1, for one child, pointed to by op_first.
struct unop {
BASEOP
OP * op_first;
}
$ perl -F"/\cI+/" -ane 'print if $F[3] =~ /1$/' opcode.pl
rv2gv ref-to-glob cast ck_rvconst ds1
rv2sv scalar dereference ck_rvconst ds1
av2arylen array length ck_null is1
rv2cv subroutine dereference ck_rvconst d1
refgen reference constructor ck_spair m1 L
srefgen single ref constructor ck_null fs1 S
regcmaybe regexp internal guard ck_fun s1 S
regcreset regexp internal reset ck_fun s1 S
preinc preincrement (++) ck_lfun dIs1 S
i_preinc integer preincrement (++) ck_lfun dis1 S
predec predecrement (--) ck_lfun dIs1 S
i_predec integer predecrement (--) ck_lfun dis1 S
postinc postincrement (++) ck_lfun dIst1 S
i_postinc integer postincrement (++) ck_lfun disT1 S
postdec postdecrement (--) ck_lfun dIst1 S
i_postdec integer postdecrement (--) ck_lfun disT1 S
negate negation (-) ck_null Ifst1 S
i_negate integer negation (-) ck_null ifsT1 S
not not ck_null ifs1 S
complement 1's complement (~) ck_bitop fst1 S
rv2av array dereference ck_rvconst dt1
rv2hv hash dereference ck_rvconst dt1
flip range (or flip) ck_null 1 S S
flop range (or flop) ck_null 1
method method lookup ck_method d1
entersub subroutine entry ck_subr dmt1 L
leavesub subroutine exit ck_null 1
leavesublv lvalue subroutine return ck_null 1
leavegiven leave given block ck_null 1
leavewhen leave when block ck_null 1
leavewrite write exit ck_null 1
dofile do "file" ck_fun d1 S
leaveeval eval "string" exit ck_null 1 S
#evalonce eval constant string ck_null d1 SBINOP
The BINOP class signifier is 2, for two children, pointed to by op_first and op_last.
struct binop {
BASEOP
OP * op_first;
OP * op_last;
}
$ perl -F"/\cI+/" -ane 'print if $F[3] =~ /2$/' opcode.pl
gelem glob elem ck_null d2 S S
aassign list assignment ck_null t2 L L
pow exponentiation (**) ck_null fsT2 S S
multiply multiplication (*) ck_null IfsT2 S S
i_multiply integer multiplication (*) ck_null ifsT2 S S
divide division (/) ck_null IfsT2 S S
i_divide integer division (/) ck_null ifsT2 S S
modulo modulus (%) ck_null IifsT2 S S
i_modulo integer modulus (%) ck_null ifsT2 S S
repeat repeat (x) ck_repeat mt2 L S
add addition (+) ck_null IfsT2 S S
i_add integer addition (+) ck_null ifsT2 S S
subtract subtraction (-) ck_null IfsT2 S S
i_subtract integer subtraction (-) ck_null ifsT2 S S
concat concatenation (.) or string ck_concat fsT2 S S
left_shift left bitshift (<<) ck_bitop fsT2 S S
right_shift right bitshift (>>) ck_bitop fsT2 S S
lt numeric lt (<) ck_null Iifs2 S S
i_lt integer lt (<) ck_null ifs2 S S
gt numeric gt (>) ck_null Iifs2 S S
i_gt integer gt (>) ck_null ifs2 S S
le numeric le (<=) ck_null Iifs2 S S
i_le integer le (<=) ck_null ifs2 S S
ge numeric ge (>=) ck_null Iifs2 S S
i_ge integer ge (>=) ck_null ifs2 S S
eq numeric eq (==) ck_null Iifs2 S S
i_eq integer eq (==) ck_null ifs2 S S
ne numeric ne (!=) ck_null Iifs2 S S
i_ne integer ne (!=) ck_null ifs2 S S
ncmp numeric comparison (<=>)ck_null Iifst2 S S
i_ncmp integer comparison (<=>)ck_null ifst2 S S
slt string lt ck_null ifs2 S S
sgt string gt ck_null ifs2 S S
sle string le ck_null ifs2 S S
sge string ge ck_null ifs2 S S
seq string eq ck_null ifs2 S S
sne string ne ck_null ifs2 S S
scmp string comparison (cmp) ck_null ifst2 S S
bit_and bitwise and (&) ck_bitop fst2 S S
bit_xor bitwise xor (^) ck_bitop fst2 S S
bit_or bitwise or (|) ck_bitop fst2 S S
smartmatch smart match ck_smartmatch s2
aelem array element ck_null s2 A S
helem hash element ck_null s2 H S
lslice list slice ck_null 2 H L L
xor logical xor ck_null fs2 S S
leaveloop loop exit ck_null 2
LOGOP
The LOGOP class signifier is |.
A LOGOP has the same structure as a "BINOP", two children, just the second field has another name op_other instead of op_last. But as you see on the list below, the two arguments as above are optional and not strictly required.
struct logop {
BASEOP
OP * op_first;
OP * op_other;
};
$ perl -F"/\cI+/" -ane 'print if $F[3] =~ /\|$/' opcode.pl
regcomp regexp compilation ck_null s| S
substcont substitution iterator ck_null dis|
grepwhile grep iterator ck_null dt|
mapwhile map iterator ck_null dt|
range flipflop ck_null | S S
and logical and (&&) ck_null |
or logical or (||) ck_null |
dor defined or (//) ck_null |
cond_expr conditional expression ck_null d|
andassign logical and assignment (&&=) ck_null s|
orassign logical or assignment (||=) ck_null s|
dorassign defined or assignment (//=) ck_null s|
entergiven given() ck_null d|
enterwhen when() ck_null d|
entertry eval {block} ck_null |
once once ck_null |LISTOP
The LISTOP class signifier is @.
struct listop {
BASEOP
OP * op_first;
OP * op_last;
};This is most complex type, it may have any number of children. The first child is pointed to by op_first and the last child by op_last. The children in between can be found by iteratively following the op_sibling pointer from the first child to the last.
At all 99 ops from 366 are LISTOP's. This is the least restrictive format, that's why.
$ perl -F"/\cI+/" -ane 'print if $F[3] =~ /\@$/' opcode.pl
bless bless ck_fun s@ S S?
glob glob ck_glob t@ S?
stringify string ck_fun fsT@ S
atan2 atan2 ck_fun fsT@ S S
substr substr ck_substr st@ S S S? S?
vec vec ck_fun ist@ S S S
index index ck_index isT@ S S S?
rindex rindex ck_index isT@ S S S?
sprintf sprintf ck_fun fmst@ S L
formline formline ck_fun ms@ S L
crypt crypt ck_fun fsT@ S S
aslice array slice ck_null m@ A L
hslice hash slice ck_null m@ H L
unpack unpack ck_unpack @ S S?
pack pack ck_fun mst@ S L
split split ck_split t@ S S S
join join or string ck_join mst@ S L
list list ck_null m@ L
anonlist anonymous list ([]) ck_fun ms@ L
anonhash anonymous hash ({}) ck_fun ms@ L
splice splice ck_fun m@ A S? S? L
... and so on, until
syscall syscall ck_fun imst@ S LPMOP
The PMOP "pattern matching" class signifier is / for matching. It inherits from the "LISTOP".
The internal struct changed completely with 5.10, as the underlying engine. Starting with 5.11 the PMOP can even hold native "perlguts#REGEX" in "REGEX" objects, not just SV's. So you have to use the PM macros to stay compatible.
Below is the current struct pmop. You will not like it.
struct pmop {
BASEOP
OP * op_first;
OP * op_last;
#ifdef USE_ITHREADS
IV op_pmoffset;
#else
REGEXP * op_pmregexp; /* compiled expression */
#endif
U32 op_pmflags;
union {
OP * op_pmreplroot; /* For OP_SUBST */
#ifdef USE_ITHREADS
PADOFFSET op_pmtargetoff; /* For OP_PUSHRE */
#else
GV * op_pmtargetgv;
#endif
} op_pmreplrootu;
union {
OP * op_pmreplstart; /* Only used in OP_SUBST */
#ifdef USE_ITHREADS
char * op_pmstashpv; /* Only used in OP_MATCH, with PMf_ONCE set */
#else
HV * op_pmstash;
#endif
} op_pmstashstartu;
};Before we had no union, but a op_pmnext, which never worked. Maybe because of the typo in the comment.
The old struct (up to 5.8.x) was as simple as:
struct pmop {
BASEOP
OP * op_first;
OP * op_last;
U32 op_children;
OP * op_pmreplroot;
OP * op_pmreplstart;
PMOP * op_pmnext; /* list of all scanpats */
REGEXP * op_pmregexp; /* compiled expression */
U16 op_pmflags;
U16 op_pmpermflags;
U8 op_pmdynflags;
}So op_pmnext, op_pmpermflags and op_pmdynflags are gone. The op_pmflags are not the whole deal, there's also op_pmregexp.extflags - interestingly called B::PMOP::reflags in B - for the new features. This is btw. the only inconsistency in the B mapping.
$ perl -F"/\cI+/" -ane 'print if $F[3] =~ /\/$/' opcode.pl
pushre push regexp ck_null d/
match pattern match (m//) ck_match d/
qr pattern quote (qr//) ck_match s/
subst substitution (s///) ck_match dis/ SSVOP
The SVOP class is very special, and can even change dynamically. Whole SV's are costly and are now just used as GV or RV. The SVOP has no special signifier, as there are different subclasses. See "SVOP_OR_PADOP", "PVOP_OR_SVOP" and "FILESTATOP".
A SVOP holds a SV and is in case of an FILESTATOP the GV for the filehandle argument, and in case of trans (a "PVOP") with utf8 a reference to a swash (i.e., an RV pointing to an HV).
struct svop {
BASEOP
SV * op_sv;
};Most old SVOP's were changed to "PADOP"'s when threading was introduced, to privatize the global SV area to thread-local scratchpads.
SVOP_OR_PADOP
The op aelemfast is either a PADOP with threading and a simple SVOP without. This is thanksfully known at compile-time.
aelemfast constant array element ck_null s$ A SPVOP_OR_SVOP
The only op here is trans, where the class is dynamically defined, dependent on the utf8 settings in the "op_private" hints.
case OA_PVOP_OR_SVOP:
return (o->op_private & (OPpTRANS_TO_UTF|OPpTRANS_FROM_UTF))
? OPc_SVOP : OPc_PVOP;
trans transliteration (tr///) ck_null is" SCharacter translations (tr///) are usually a PVOP, keeping a pointer to a table of shorts used to look up translations. Under utf8, however, a simple table isn't practical; instead, the OP is an "SVOP", and the SV is a reference to a swash, i.e. a RV pointing to an HV.
PADOP
The PADOP class signifier is $ for temp. scalars.
A new PADOP creates a new temporary scratchpad, an PADLIST array. padop-op_padix = pad_alloc(type, SVs_PADTMP);> SVs_PADTMP are targets/GVs/constants with undef names.
A PADLIST scratchpad is a special context stack, a array-of-array data structure attached to a CV (i.e. a sub), to store lexical variables and opcode temporary and per-thread values. See "Scratchpads" in perlguts.
Only my/our variable (SVs_PADMY/SVs_PADOUR) slots get valid names. The rest are op targets/GVs/constants which are statically allocated or resolved at compile time. These don't have names by which they can be looked up from Perl code at run time through eval "" like my/our variables can be. Since they can't be looked up by "name" but only by their index allocated at compile time (which is usually in op_targ), wasting a name SV for them doesn't make sense.
struct padop {
BASEOP
PADOFFSET op_padix;
};
$ perl -F"/\cI+/" -ane 'print if $F[3] =~ /\$$/' opcode.pl
const constant item ck_svconst s$
gvsv scalar variable ck_null ds$
gv glob value ck_null ds$
anoncode anonymous subroutine ck_anoncode $
rcatline append I/O operator ck_null t$
aelemfast constant array element ck_null s$ A S
method_named method with known name ck_null d$
hintseval eval hints ck_svconst s$PVOP
This is a simple unary op, holding a string. The only PVOP is trans op for "//" in tr. See above at "PVOP_OR_SVOP" for the dynamic nature of trans with utf8.
The PVOP class signifier is " for strings.
struct pvop {
BASEOP
char * op_pv;
};
$ perl -F"/\cI+/" -ane 'print if $F[3] =~ /\"$/' opcode.pl
trans transliteration (tr///) ck_match is" SLOOP
The LOOP class signifier is {. It inherits from the "LISTOP".
struct loop {
BASEOP
OP * op_first;
OP * op_last;
OP * op_redoop;
OP * op_nextop;
OP * op_lastop;
};
$ perl -F"/\cI+/" -ane 'print if $F[3] =~ /\{$/' opcode.pl
enteriter foreach loop entry ck_null d{
enterloop loop entry ck_null d{COP
The struct cop, the "Control OP", changed recently a lot, as the "BASEOP". Remember from perlguts what a COP is? Got you. A COP is nowhere described.
I would have naively called it "Context OP", but not "Control OP". So why? We have a global PL_curcop and then we have threads. So it cannot be global anymore. A COP can be said as helper context for debugging and error information to store away file and line information. But since perl is a file-based compiler, not block-based, also file based pragmata and hints are stored in the COP. So we have for every source file a seperate COP. COP's are mostly not really block level contexts, just file and line information. The block level contexts are not controlled via COP's, but global Cx structs.
cop.h says:
Control ops (cops) are one of the three ops OP_NEXTSTATE, OP_DBSTATE, and OP_SETSTATE that (loosely speaking) are separate statements. They hold information for lexical state and error reporting. At run time, PL_curcop is set to point to the most recently executed cop, and thus can be used to determine our file-level current state.
But we need block context, eval context, subroutine context, loop context, and even format context. All these are seperate structs defined in cop.h.
So the COPs are not really that important, as the actual Cx context structs are. Just the CopSTASH is, the current package symbol table hash ("stash").
Another famous COP is PL_compiling, which sets the temporary compilation environment.
struct cop {
BASEOP
line_t cop_line; /* line # of this command */
char * cop_label; /* label for this construct */
#ifdef USE_ITHREADS
char * cop_stashpv; /* package line was compiled in */
char * cop_file; /* file name the following line # is from */
#else
HV * cop_stash; /* package line was compiled in */
GV * cop_filegv; /* file the following line # is from */
#endif
U32 cop_hints; /* hints bits from pragmata */
U32 cop_seq; /* parse sequence number */
/* Beware. mg.c and warnings.pl assume the type of this is STRLEN *: */
STRLEN * cop_warnings; /* lexical warnings bitmask */
/* compile time state of %^H. See the comment in op.c for how this is
used to recreate a hash to return from caller. */
struct refcounted_he * cop_hints_hash;
};The COP class signifier is ; and there are only two:
$ perl -F"/\cI+/" -ane 'print if $F[3] =~ /;$/' opcode.pl
nextstate next statement ck_null s;
dbstate debug next statement ck_null s;NEXTSTATE is replaced by DBSTATE when you call perl with -d, the debugger. You can even patch the NEXTSTATE ops at runtime to DBSTATE as done in the module Enbugger.
For a short time there used to be three. SETSTATE was added 1999 (pre Perl 5.6.0) to track linenumbers correctly in optimized blocks, disabled 1999 with change 4309 for Perl 5.6.0, and removed with 5edb5b2abb at Perl 5.10.1.
This requires a patch to remove wrong info from the cop.h header. Read perlhack.pod and submit the patch after investigating where it has gone to. Sam Vilain's GIT history might help.
BASEOP_OR_UNOP
BASEOP_OR_UNOP has the class signifier %. As the name says, it may be a "BASEOP" or "UNOP", it may have an optional "op_first" field.
The list of % ops is quite large, it has 84 ops. Some of them are e.g.
$ perl -F"/\cI+/" -ane 'print if $F[3] =~ /%$/' opcode.pl
...
quotemeta quotemeta ck_fun fstu% S?
aeach each on array ck_each % A
akeys keys on array ck_each t% A
avalues values on array ck_each t% A
each each ck_each % H
values values ck_each t% H
keys keys ck_each t% H
delete delete ck_delete % S
exists exists ck_exists is% S
pop pop ck_shift s% A?
shift shift ck_shift s% A?
caller caller ck_fun t% S?
reset symbol reset ck_fun is% S?
exit exit ck_exit ds% S?
...FILESTATOP
A FILESTATOP may be a "UNOP", "PADOP", "BASEOP" or "SVOP".
It has the class signifier -.
The file stat OPs are created via UNI(OP_foo) in toke.c but use the OPf_REF flag to distinguish between OP types instead of the usual OPf_SPECIAL flag. As usual, if OPf_KIDS is set, then we return OPc_UNOP so that walkoptree can find our children. If OPf_KIDS is not set then we check OPf_REF. Without OPf_REF set (no argument to the operator) it's an OP; with OPf_REF set it's an SVOP (and the field op_sv is the GV for the filehandle argument).
case OA_FILESTATOP:
return ((o->op_flags & OPf_KIDS) ? OPc_UNOP :
#ifdef USE_ITHREADS
(o->op_flags & OPf_REF) ? OPc_PADOP : OPc_BASEOP);
#else
(o->op_flags & OPf_REF) ? OPc_SVOP : OPc_BASEOP);
#endif
lstat lstat ck_ftst u- F
stat stat ck_ftst u- F
ftrread -R ck_ftst isu- F-+
ftrwrite -W ck_ftst isu- F-+
ftrexec -X ck_ftst isu- F-+
fteread -r ck_ftst isu- F-+
ftewrite -w ck_ftst isu- F-+
fteexec -x ck_ftst isu- F-+
ftis -e ck_ftst isu- F-
ftsize -s ck_ftst istu- F-
ftmtime -M ck_ftst stu- F-
ftatime -A ck_ftst stu- F-
ftctime -C ck_ftst stu- F-
ftrowned -O ck_ftst isu- F-
fteowned -o ck_ftst isu- F-
ftzero -z ck_ftst isu- F-
ftsock -S ck_ftst isu- F-
ftchr -c ck_ftst isu- F-
ftblk -b ck_ftst isu- F-
ftfile -f ck_ftst isu- F-
ftdir -d ck_ftst isu- F-
ftpipe -p ck_ftst isu- F-
ftsuid -u ck_ftst isu- F-
ftsgid -g ck_ftst isu- F-
ftsvtx -k ck_ftst isu- F-
ftlink -l ck_ftst isu- F-
fttty -t ck_ftst is- F-
fttext -T ck_ftst isu- F-
ftbinary -B ck_ftst isu- F-LOOPEXOP
A LOOPEXOP is almost a BASEOP_OR_UNOP. It may be a "UNOP" if stacked or "BASEOP" if special or "PVOP" else.
next, last, redo, dump and goto use OPf_SPECIAL to indicate that a label was omitted (in which case it's a "BASEOP") or else a term was seen. In this last case, all except goto are definitely "PVOP" but goto is either a PVOP (with an ordinary constant label), an "UNOP" with OPf_STACKED (with a non-constant non-sub) or an "UNOP" for OP_REFGEN (with goto &sub) in which case OPf_STACKED also seems to get set.
...
OP Definition Example
Let's take a simple example for a opcode definition in opcode.pl:
left_shift left bitshift (<<) ck_bitop fsT2 S SThe op left_shift has a check function ck_bitop (normally most ops have no check function, just ck_null), and the options fsT2. The last two S S describe the type of the two required operands: SV or scalar. This is similar to XS protoypes. The last 2 in the options fsT2 denotes the class BINOP, with two args on the stack. Every binop takes two args and this produces one scalar, see the s flag. The other remaining flags are f and T.
f tells the compiler in the first pass to call fold_constants() on this op. See "Compile pass 1: check routines and constant folding" If both args are constant, the result is constant also and the op will be nullified.
Now let's inspect the simple definition of this op in pp.c. pp_left_shift is the op_ppaddr, the function pointer, for every left_shift op.
PP(pp_left_shift)
{
dVAR; dSP; dATARGET; tryAMAGICbin(lshift,opASSIGN);
{
const IV shift = POPi;
if (PL_op->op_private & HINT_INTEGER) {
const IV i = TOPi;
SETi(i << shift);
}
else {
const UV u = TOPu;
SETu(u << shift);
}
RETURN;
}
}The first IV arg is pop'ed from the stack, the second arg is left on the stack (TOPi/TOPu), because it is used as the return value. (Todo: explain the opASSIGN magic check.) One IV or UV is produced, dependent on HINT_INTEGER, set by the use integer pragma. So it has a special signed/unsigned integer behaviour, which is not defined in the opcode declaration, because the API is indifferent on this, and it is also independent on the argument type. The result, if IV or UV, is entirely context dependent at compile-time ( use integer at BEGIN ) or run-time ( $^H |= 1 ), and only stored in the op.
What is left is the T flag, "target can be a pad". This is a useful optimization technique.
This is checked in the macro dATARGET SV *targ = (PL_op-op_flags & OPf_STACKED ? sp[-1] : PAD_SV(PL_op->op_targ));> OPf_STACKED means "Some arg is arriving on the stack." (see op.h) So this reads, if the op contains OPf_STACKED, the magic targ ("target argument") is simply on the stack, but if not, the op_targ points to a SV on a private scratchpad. "target can be a pad", voila. For reference see "Putting a C value on Perl stack" in perlguts.
Check Functions
They are defined in op.c and not in pp.c, because they belong tightly to the ops and newOP definition, and not to the actual pp_ opcode. That's why the actual op.c file is bigger than pp.c where the real gore for each op begins. The name of each op's check function is defined in opcodes.pl, as shown above.
The ck_null check function is the most common.
$ perl -F"/\cI+/" -ane 'print $F[2],"\n" if $F[2] =~ /ck_null/' opcode.pl|wc -l
128But we do have a lot of those check functions.
$ perl -F"/\cI+/" -ane 'print $F[2],"\n" if $F[2] =~ /ck_/' opcode.pl|sort -u|wc -l
43When are they called, how do they look like, what do they do.
The macro CHECKOP(type,o) used to call the ck_ function has a little bit of common logic.
#define CHECKOP(type,o) \
((PL_op_mask && PL_op_mask[type]) \
? ( op_free((OP*)o), \
Perl_croak(aTHX_ "'%s' trapped by operation mask", PL_op_desc[type]), \
(OP*)0 ) \
: CALL_FPTR(PL_check[type])(aTHX_ (OP*)o))So when a global PL_op_mask is fitting to the type the OP is nullified at once. If not the type specific check function, with the help of opcodes.pl generating the PL_check array in opnames.h is called.
Constant Folding
In theory pretty easy. If all op's arguments are constant and the op is sideffect free ("purely functional"), replace the op with an constant op as result.
We do it like this: We define the f flag in opcodes.pl, which tells the compiler in the first pass to call fold_constants() on this op. See "Compile pass 1: check routines and constant folding" above. If all args are constant, the result is constant also and the op will be nullified.
But take care, every f op must be sideeffect free!
E.g. our newUNOP() calls at the end:
return fold_constants((OP *) unop);OA_FOLDCONST
Lexical Pragmas
To implement user lexical pragmas, there needs to be a way at run time to get the compile time state of `%^H` for that block. Storing `%^H` in every block (or even COP) would be very expensive, so a different approach is taken. The (running) state of %^H is serialised into a tree of HE-like structs. Stores into %^H are chained onto the current leaf as a struct refcounted_he * with the key and the value. Deletes from %^H are saved with a value of PL_sv_placeholder. The state of %^H at any point can be turned back into a regular HV by walking back up the tree from that point's leaf, ignoring any key you've already seen (placeholder or not), storing the rest into the HV structure, then removing the placeholders. Hence memory is only used to store the %^H deltas from the enclosing COP, rather than the entire %^H on each COP.
To cause actions on %^H to write out the serialisation records, it has magic type 'H'. This magic (itself) does nothing, but its presence causes the values to gain magic type 'h', which has entries for set and clear. Perl_magic_sethint updates PL_compiling.cop_hints_hash with a store record, with deletes written by Perl_magic_clearhint. SAVEHINTS saves the current PL_compiling.cop_hints_hash on the save stack, so that it will be correctly restored when any inner compiling scope is exited.
Hooks
Special execution blocks BEGIN, CHECK, UNITCHECK, INIT, END
Perl keeps special arrays of subroutines that are executed at the beginning and at the end of a running Perl program and its program units. These subroutines correspond to the special code blocks: BEGIN, CHECK, UNITCHECK, INIT and END. (See basics at "basics" in perlmod.)
Such arrays belong to Perl's internals that you're not supposed to see. Entries in these arrays get consumed by the interpreter as it enters distinct compilation phases, triggered by statements like require, use, do, eval, etc. To play as safest as possible, the only allowed operations are to add entries to the start and to the end of these arrays.
BEGIN, UNITCHECK and INIT are FIFO (first-in, first-out) blocks while CHECK and END are LIFO (last-in, first-out).
Devel::Hook allows adding code the start or end of these blocks. Manip::END even tries to remove certain entries.
The BEGIN block
A special array of code at PL_beginav, that is executed before main_start, the first op, which is defined be called ENTER. E.g. use module; adds its require and importer code into the BEGIN block.
The CHECK block
The B compiler starting block at PL_checkav. This hooks int the check function which is executed for every op created in bottom-up, basic order.
The UNITCHECK block
A new block since Perl 5.10 at PL_unitcheckav runs right after the CHECK block, to seperate possible B compilation hooks from other checks.
The INIT block
At PL_initav.
The END block
At PL_endav.
Manip::END started to mess around with this block.
The array contains an undef for each block that has been encountered. It's not really an undef though, it's a kind of raw coderef that's not wrapped in a scalar ref. This leads to funky error messages like Bizarre copy of CODE in sassign when you try to assign one of these values to another variable. See Manip::END how to manipulate these values array.
B and O module. The perl compiler.
Malcom Beattie's B modules hooked into the early op tree stages to represent the internal ops as perl objects and added the perl compiler backends. See B and perlcompile.pod.
The three main compiler backends are still Bytecode, C and CC, and these do not work for current perl's (yet).
Todo: Describe B's object representation a little bit deeper, its CHECK hook, its internal transformers for Bytecode (asm and vars) and C (the sections).
MAD
MAD stands for "Misc Attributed Data".
Larry Wall worked on a new MAD compiler backend outside of the B approach, dumping the internal op tree representation as XML, not as tree of perl B objects.
The idea is that all the information needed to recreate the original source is stored in the op tree. To do this the tokens for the ops are associated with ops, these madprops are a list of key-value pairs, where the key is a character as listed at the end of op.h, the value normally is a string, but it might also be a op, as in the case of a optimized op ('O'). Special for the whitespace key '_' (whitespace before) and '#' (whitespace after), which indicate the whitespace or comment before/after the previous key-value pair.
Also when things normally compiled out, like a BEGIN block, which normally do not results in any ops, instead create a NULLOP with madprops used to recreate the object.
Is there any documentation on this?
Why this awful XML and not the rich tree of perl objects?
Well there's an advantage. The MAD XML can be seen as some kind of XML Storable/Freeze of the B op tree, and can be therefore converted outside of the CHECK block, which means you can actually debug the conversion (= compilation) process. This is not possible within the CHECK block in the B backends.
kurila http://search.cpan.org/dist/kurila/ uses this to convert Perl 5 source to the kurila dialect.
To convert a file 'source.pm' from Perl 5.10 to Kurila you need to do:
kurilapath=/usr/src/perl/kurila-1.9
bleadpath=/usr/src/perl/blead
cd $kurilapath
madfrom='perl-5.10' madto='kurila-1.9' \
madconvert="/usr/bin/perl $kurilapath/mad/p5kurila.pl" \
madpath="$bleadpath/mad" \
mad/convert /path/to/source.pmPPI http://search.cpan.org/dist/PPI/, a Perl 5 source level parser not related to the op tree at all, could also have been used for that.
Pluggable runops
The compile tree is executed by one of two existing runops functions, in run.c or in dump.c. Perl_runops_debug is used with DEBUGGING and the faster Perl_runops_standard is used otherwise (See below in "Walkers"). For fine control over the execution of the compile tree it is possible to provide your own runops function.
It's probably best to copy one of the existing runops functions and change it to suit your needs. Then, in the BOOT section of your XS file, add the line:
PL_runops = my_runops;This function should be as efficient as possible to keep your programs running as fast as possible.
Walkers or runops
The standard op tree walker or runops is as simple as this ultrafast Perl_runops_standard() in (run.c). It starts with main_start and walks the op_next chain until the end. No need to check other fields, strictly linear through the tree.
int
Perl_runops_standard(pTHX)
{
dVAR;
while ((PL_op = CALL_FPTR(PL_op->op_ppaddr)(aTHX))) {
PERL_ASYNC_CHECK();
}
TAINT_NOT;
return 0;
}To inspect the op tree within a perl program, you can also hook PL_runops (see above at "Pluggable runops") to your own perl walker (see e.g. B::Utils for various useful walkers), but you cannot modify the tree from within the B accessors, only via XS. Or via B::Generate as explained in Simon Cozen's "Hacking the Optree for Fun..." http://www.perl.com/pub/a/2002/05/07/optree.html.
Todo: Show the other runloops, and esp. the B:Utils ones. Todo: Describe the dumper, the debugging and more extended walkers.
Internal and external modifications
See the short description of the internal optimizer in the "Brief Summary".
Todo: Describe the exported variables and functions which can be hooked, besides simply adding code to the blocks.
Via "Pluggable runops" you can provide your own walker function, as it is done in most B modules. Best see B::Utils.
You may also create custom ops at runtime (well, strictly speaking at compile-time) via B::Generate.
Modules
The most important op tree module is B::Concise by Stephen McCamant.
B::Utils provides abstract-enough op tree grep's and walkers with callbacks from the perl level.
Devel::Hook allows adding perl hooks into the BEGIN, CHECK, UNITCHECK, INIT blocks.
Devel::TypeCheck tries to verify possible static typing for expressions and variables, a pretty hard problem for compilers, esp. with such dynamic and untyped variables as Perl 5.
Reini Urban is working on an interactive op tree debugger, B::Debugger.
Various Articles
The best source of information is the source. It is very well documented.
Simon Cozens has posted the course material to NetThink's http://books.simon-cozens.org/index.php/Perl_5_Internals#The_Lexer_and_the_Parser training course. This is the currently best available description on that subject.
"Hacking the Optree for Fun..." at http://www.perl.com/pub/a/2002/05/07/optree.html is the next step by Simon Cozens.
Joshua ben Jore wrote a 50 minute presentation on "Perl 5 VM guts" at http://diotalevi.isa-geek.net/~josh/Presentations/Perl%205%20VM/ focusing on the op tree for SPUG, the Seattle Perl User's Group.
Eric Wilhelm wrote a brief tour through the perl compiler backends for the impatient refactorerer. The perl_guts_tour as mp3 http://scratchcomputing.com/developers/perl_guts_tour.html or as pdf http://scratchcomputing.com/developers/perl_guts_tour.pdf
This text was created in this wiki article: http://www.perlfoundation.org/perl5/index.cgi?optree_guts The wiki article should be more actual.
Conclusion
So this is about 30% of the basic op tree information so far. Not speaking about the guts. Simon has more 30%, in the source are more 10% to copy&paste, and in the compilers and run-time information is the rest. I hope with the help of some hackers we'll get it done, so that some people will begin poking around in the B backends. And write the wonderful new dump/undump functionality (which actually worked in the early years on Solaris), analyse and optimize the output, output PIR (parrot code), emit LLVM or another JIT optimized code or even write assemblers. I have a simple one at home. :)
Written 2008 on the perl5 wiki with socialtext and pod in parallel by Reini Urban, CPAN ID rurban.