2011-09-15 Tom de Vries <tom@codesourcery.com>

[official-gcc.git] / gcc / doc / gty.texi

@c Copyright (C) 2002, 2003, 2004, 2007, 2008, 2009, 2010
@c Free Software Foundation, Inc.
@c This is part of the GCC manual.
@c For copying conditions, see the file gcc.texi.

@node Type Information
@chapter Memory Management and Type Information
@cindex GGC
@findex GTY

GCC uses some fairly sophisticated memory management techniques, which
involve determining information about GCC's data structures from GCC's
source code and using this information to perform garbage collection and
implement precompiled headers.

A full C parser would be too complicated for this task, so a limited
subset of C is interpreted and special markers are used to determine
what parts of the source to look at.  All @code{struct} and
@code{union} declarations that define data structures that are
allocated under control of the garbage collector must be marked.  All
global variables that hold pointers to garbage-collected memory must
also be marked.  Finally, all global variables that need to be saved
and restored by a precompiled header must be marked.  (The precompiled
header mechanism can only save static variables if they're scalar.
Complex data structures must be allocated in garbage-collected memory
to be saved in a precompiled header.)

The full format of a marker is
@smallexample
GTY (([@var{option}] [(@var{param})], [@var{option}] [(@var{param})] @dots{}))
@end smallexample
@noindent
but in most cases no options are needed.  The outer double parentheses
are still necessary, though: @code{GTY(())}.  Markers can appear:

@itemize @bullet
@item
In a structure definition, before the open brace;
@item
In a global variable declaration, after the keyword @code{static} or
@code{extern}; and
@item
In a structure field definition, before the name of the field.
@end itemize

Here are some examples of marking simple data structures and globals.

@smallexample
struct GTY(()) @var{tag}
@{
  @var{fields}@dots{}
@};

typedef struct GTY(()) @var{tag}
@{
  @var{fields}@dots{}
@} *@var{typename};

static GTY(()) struct @var{tag} *@var{list};   /* @r{points to GC memory} */
static GTY(()) int @var{counter};        /* @r{save counter in a PCH} */
@end smallexample

The parser understands simple typedefs such as
@code{typedef struct @var{tag} *@var{name};} and
@code{typedef int @var{name};}.
These don't need to be marked.

@menu
* GTY Options::         What goes inside a @code{GTY(())}.
* GGC Roots::           Making global variables GGC roots.
* Files::               How the generated files work.
* Invoking the garbage collector::   How to invoke the garbage collector.
* Troubleshooting::     When something does not work as expected.
@end menu

@node GTY Options
@section The Inside of a @code{GTY(())}

Sometimes the C code is not enough to fully describe the type
structure.  Extra information can be provided with @code{GTY} options
and additional markers.  Some options take a parameter, which may be
either a string or a type name, depending on the parameter.  If an
option takes no parameter, it is acceptable either to omit the
parameter entirely, or to provide an empty string as a parameter.  For
example, @code{@w{GTY ((skip))}} and @code{@w{GTY ((skip ("")))}} are
equivalent.

When the parameter is a string, often it is a fragment of C code.  Four
special escapes may be used in these strings, to refer to pieces of
the data structure being marked:

@cindex % in GTY option
@table @code
@item %h
The current structure.
@item %1
The structure that immediately contains the current structure.
@item %0
The outermost structure that contains the current structure.
@item %a
A partial expression of the form @code{[i1][i2]@dots{}} that indexes
the array item currently being marked.
@end table

For instance, suppose that you have a structure of the form
@smallexample
struct A @{
  @dots{}
@};
struct B @{
  struct A foo[12];
@};
@end smallexample
@noindent
and @code{b} is a variable of type @code{struct B}.  When marking
@samp{b.foo[11]}, @code{%h} would expand to @samp{b.foo[11]},
@code{%0} and @code{%1} would both expand to @samp{b}, and @code{%a}
would expand to @samp{[11]}.

As in ordinary C, adjacent strings will be concatenated; this is
helpful when you have a complicated expression.
@smallexample
@group
GTY ((chain_next ("TREE_CODE (&%h.generic) == INTEGER_TYPE"
                  " ? TYPE_NEXT_VARIANT (&%h.generic)"
                  " : TREE_CHAIN (&%h.generic)")))
@end group
@end smallexample

The available options are:

@table @code
@findex length
@item length ("@var{expression}")

There are two places the type machinery will need to be explicitly told
the length of an array.  The first case is when a structure ends in a
variable-length array, like this:
@smallexample
struct GTY(()) rtvec_def @{
  int num_elem;         /* @r{number of elements} */
  rtx GTY ((length ("%h.num_elem"))) elem[1];
@};
@end smallexample

In this case, the @code{length} option is used to override the specified
array length (which should usually be @code{1}).  The parameter of the
option is a fragment of C code that calculates the length.

The second case is when a structure or a global variable contains a
pointer to an array, like this:
@smallexample
struct gimple_omp_for_iter * GTY((length ("%h.collapse"))) iter;
@end smallexample
In this case, @code{iter} has been allocated by writing something like
@smallexample
  x->iter = ggc_alloc_cleared_vec_gimple_omp_for_iter (collapse);
@end smallexample
and the @code{collapse} provides the length of the field.

This second use of @code{length} also works on global variables, like:
@verbatim
static GTY((length("reg_known_value_size"))) rtx *reg_known_value;
@end verbatim

@findex skip
@item skip

If @code{skip} is applied to a field, the type machinery will ignore it.
This is somewhat dangerous; the only safe use is in a union when one
field really isn't ever used.

@findex desc
@findex tag
@findex default
@item desc ("@var{expression}")
@itemx tag ("@var{constant}")
@itemx default

The type machinery needs to be told which field of a @code{union} is
currently active.  This is done by giving each field a constant
@code{tag} value, and then specifying a discriminator using @code{desc}.
The value of the expression given by @code{desc} is compared against
each @code{tag} value, each of which should be different.  If no
@code{tag} is matched, the field marked with @code{default} is used if
there is one, otherwise no field in the union will be marked.

In the @code{desc} option, the ``current structure'' is the union that
it discriminates.  Use @code{%1} to mean the structure containing it.
There are no escapes available to the @code{tag} option, since it is a
constant.

For example,
@smallexample
struct GTY(()) tree_binding
@{
  struct tree_common common;
  union tree_binding_u @{
    tree GTY ((tag ("0"))) scope;
    struct cp_binding_level * GTY ((tag ("1"))) level;
  @} GTY ((desc ("BINDING_HAS_LEVEL_P ((tree)&%0)"))) xscope;
  tree value;
@};
@end smallexample

In this example, the value of BINDING_HAS_LEVEL_P when applied to a
@code{struct tree_binding *} is presumed to be 0 or 1.  If 1, the type
mechanism will treat the field @code{level} as being present and if 0,
will treat the field @code{scope} as being present.

@findex param_is
@findex use_param
@item param_is (@var{type})
@itemx use_param

Sometimes it's convenient to define some data structure to work on
generic pointers (that is, @code{PTR}) and then use it with a specific
type.  @code{param_is} specifies the real type pointed to, and
@code{use_param} says where in the generic data structure that type
should be put.

For instance, to have a @code{htab_t} that points to trees, one would
write the definition of @code{htab_t} like this:
@smallexample
typedef struct GTY(()) @{
  @dots{}
  void ** GTY ((use_param, @dots{})) entries;
  @dots{}
@} htab_t;
@end smallexample
and then declare variables like this:
@smallexample
  static htab_t GTY ((param_is (union tree_node))) ict;
@end smallexample

@findex param@var{n}_is
@findex use_param@var{n}
@item param@var{n}_is (@var{type})
@itemx use_param@var{n}

In more complicated cases, the data structure might need to work on
several different types, which might not necessarily all be pointers.
For this, @code{param1_is} through @code{param9_is} may be used to
specify the real type of a field identified by @code{use_param1} through
@code{use_param9}.

@findex use_params
@item use_params

When a structure contains another structure that is parameterized,
there's no need to do anything special, the inner structure inherits the
parameters of the outer one.  When a structure contains a pointer to a
parameterized structure, the type machinery won't automatically detect
this (it could, it just doesn't yet), so it's necessary to tell it that
the pointed-to structure should use the same parameters as the outer
structure.  This is done by marking the pointer with the
@code{use_params} option.

@findex deletable
@item deletable

@code{deletable}, when applied to a global variable, indicates that when
garbage collection runs, there's no need to mark anything pointed to
by this variable, it can just be set to @code{NULL} instead.  This is used
to keep a list of free structures around for re-use.

@findex if_marked
@item if_marked ("@var{expression}")

Suppose you want some kinds of object to be unique, and so you put them
in a hash table.  If garbage collection marks the hash table, these
objects will never be freed, even if the last other reference to them
goes away.  GGC has special handling to deal with this: if you use the
@code{if_marked} option on a global hash table, GGC will call the
routine whose name is the parameter to the option on each hash table
entry.  If the routine returns nonzero, the hash table entry will
be marked as usual.  If the routine returns zero, the hash table entry
will be deleted.

The routine @code{ggc_marked_p} can be used to determine if an element
has been marked already; in fact, the usual case is to use
@code{if_marked ("ggc_marked_p")}.

@findex mark_hook
@item mark_hook ("@var{hook-routine-name}")

If provided for a structure or union type, the given
@var{hook-routine-name} (between double-quotes) is the name of a
routine called when the garbage collector has just marked the data as
reachable. This routine should not change the data, or call any ggc
routine. Its only argument is a pointer to the just marked (const)
structure or union.

@findex maybe_undef
@item maybe_undef

When applied to a field, @code{maybe_undef} indicates that it's OK if
the structure that this fields points to is never defined, so long as
this field is always @code{NULL}.  This is used to avoid requiring
backends to define certain optional structures.  It doesn't work with
language frontends.

@findex nested_ptr
@item nested_ptr (@var{type}, "@var{to expression}", "@var{from expression}")

The type machinery expects all pointers to point to the start of an
object.  Sometimes for abstraction purposes it's convenient to have
a pointer which points inside an object.  So long as it's possible to
convert the original object to and from the pointer, such pointers
can still be used.  @var{type} is the type of the original object,
the @var{to expression} returns the pointer given the original object,
and the @var{from expression} returns the original object given
the pointer.  The pointer will be available using the @code{%h}
escape.

@findex chain_next
@findex chain_prev
@findex chain_circular
@item chain_next ("@var{expression}")
@itemx chain_prev ("@var{expression}")
@itemx chain_circular ("@var{expression}")

It's helpful for the type machinery to know if objects are often
chained together in long lists; this lets it generate code that uses
less stack space by iterating along the list instead of recursing down
it.  @code{chain_next} is an expression for the next item in the list,
@code{chain_prev} is an expression for the previous item.  For singly
linked lists, use only @code{chain_next}; for doubly linked lists, use
both.  The machinery requires that taking the next item of the
previous item gives the original item.  @code{chain_circular} is similar
to @code{chain_next}, but can be used for circular single linked lists.

@findex reorder
@item reorder ("@var{function name}")

Some data structures depend on the relative ordering of pointers.  If
the precompiled header machinery needs to change that ordering, it
will call the function referenced by the @code{reorder} option, before
changing the pointers in the object that's pointed to by the field the
option applies to.  The function must take four arguments, with the
signature @samp{@w{void *, void *, gt_pointer_operator, void *}}.
The first parameter is a pointer to the structure that contains the
object being updated, or the object itself if there is no containing
structure.  The second parameter is a cookie that should be ignored.
The third parameter is a routine that, given a pointer, will update it
to its correct new value.  The fourth parameter is a cookie that must
be passed to the second parameter.

PCH cannot handle data structures that depend on the absolute values
of pointers.  @code{reorder} functions can be expensive.  When
possible, it is better to depend on properties of the data, like an ID
number or the hash of a string instead.

@findex variable_size
@item variable_size

The type machinery expects the types to be of constant size.  When this
is not true, for example, with structs that have array fields or unions,
the type machinery cannot tell how many bytes need to be allocated at
each allocation.  The @code{variable_size} is used to mark such types.
The type machinery then provides allocators that take a parameter
indicating an exact size of object being allocated.  Note that the size
must be provided in bytes whereas the @code{length} option works with
array lengths in number of elements.

For example,
@smallexample
struct GTY((variable_size)) sorted_fields_type @{
  int len;
  tree GTY((length ("%h.len"))) elts[1];
@};
@end smallexample

Then the objects of @code{struct sorted_fields_type} are allocated in GC
memory as follows:
@smallexample
  field_vec = ggc_alloc_sorted_fields_type (size);
@end smallexample

If @var{field_vec->elts} stores @var{n} elements, then @var{size}
could be calculated as follows:
@smallexample
  size_t size = sizeof (struct sorted_fields_type) + n * sizeof (tree);
@end smallexample

@findex atomic
@item atomic

The @code{atomic} option can only be used with pointers.  It informs
the GC machinery that the memory that the pointer points to does not
contain any pointers, and hence it should be treated by the GC and PCH
machinery as an ``atomic'' block of memory that does not need to be
examined when scanning memory for pointers.  In particular, the
machinery will not scan that memory for pointers to mark them as
reachable (when marking pointers for GC) or to relocate them (when
writing a PCH file).

The @code{atomic} option differs from the @code{skip} option.
@code{atomic} keeps the memory under Garbage Collection, but makes the
GC ignore the contents of the memory.  @code{skip} is more drastic in
that it causes the pointer and the memory to be completely ignored by
the Garbage Collector.  So, memory marked as @code{atomic} is
automatically freed when no longer reachable, while memory marked as
@code{skip} is not.

The @code{atomic} option must be used with great care, because all
sorts of problem can occur if used incorrectly, that is, if the memory
the pointer points to does actually contain a pointer.

Here is an example of how to use it:
@smallexample
struct GTY(()) my_struct @{
  int number_of_elements;
  unsigned int GTY ((atomic)) * elements;
@};
@end smallexample
In this case, @code{elements} is a pointer under GC, and the memory it
points to needs to be allocated using the Garbage Collector, and will
be freed automatically by the Garbage Collector when it is no longer
referenced.  But the memory that the pointer points to is an array of
@code{unsigned int} elements, and the GC must not try to scan it to
find pointers to mark or relocate, which is why it is marked with the
@code{atomic} option.

Note that, currently, global variables can not be marked with
@code{atomic}; only fields of a struct can.  This is a known
limitation.  It would be useful to be able to mark global pointers
with @code{atomic} to make the PCH machinery aware of them so that
they are saved and restored correctly to PCH files.

@findex special
@item special ("@var{name}")

The @code{special} option is used to mark types that have to be dealt
with by special case machinery.  The parameter is the name of the
special case.  See @file{gengtype.c} for further details.  Avoid
adding new special cases unless there is no other alternative.
@end table

@node GGC Roots
@section Marking Roots for the Garbage Collector
@cindex roots, marking
@cindex marking roots

In addition to keeping track of types, the type machinery also locates
the global variables (@dfn{roots}) that the garbage collector starts
at.  Roots must be declared using one of the following syntaxes:

@itemize @bullet
@item
@code{extern GTY(([@var{options}])) @var{type} @var{name};}
@item
@code{static GTY(([@var{options}])) @var{type} @var{name};}
@end itemize
@noindent
The syntax
@itemize @bullet
@item
@code{GTY(([@var{options}])) @var{type} @var{name};}
@end itemize
@noindent
is @emph{not} accepted.  There should be an @code{extern} declaration
of such a variable in a header somewhere---mark that, not the
definition.  Or, if the variable is only used in one file, make it
@code{static}.

@node Files
@section Source Files Containing Type Information
@cindex generated files
@cindex files, generated

Whenever you add @code{GTY} markers to a source file that previously
had none, or create a new source file containing @code{GTY} markers,
there are three things you need to do:

@enumerate
@item
You need to add the file to the list of source files the type
machinery scans.  There are four cases:

@enumerate a
@item
For a back-end file, this is usually done
automatically; if not, you should add it to @code{target_gtfiles} in
the appropriate port's entries in @file{config.gcc}.

@item
For files shared by all front ends, add the filename to the
@code{GTFILES} variable in @file{Makefile.in}.

@item
For files that are part of one front end, add the filename to the
@code{gtfiles} variable defined in the appropriate
@file{config-lang.in}.  For C, the file is @file{c-config-lang.in}.
Headers should appear before non-headers in this list.

@item
For files that are part of some but not all front ends, add the
filename to the @code{gtfiles} variable of @emph{all} the front ends
that use it.
@end enumerate

@item
If the file was a header file, you'll need to check that it's included
in the right place to be visible to the generated files.  For a back-end
header file, this should be done automatically.  For a front-end header
file, it needs to be included by the same file that includes
@file{gtype-@var{lang}.h}.  For other header files, it needs to be
included in @file{gtype-desc.c}, which is a generated file, so add it to
@code{ifiles} in @code{open_base_file} in @file{gengtype.c}.

For source files that aren't header files, the machinery will generate a
header file that should be included in the source file you just changed.
The file will be called @file{gt-@var{path}.h} where @var{path} is the
pathname relative to the @file{gcc} directory with slashes replaced by
@verb{|-|}, so for example the header file to be included in
@file{cp/parser.c} is called @file{gt-cp-parser.c}.  The
generated header file should be included after everything else in the
source file.  Don't forget to mention this file as a dependency in the
@file{Makefile}!

@end enumerate

For language frontends, there is another file that needs to be included
somewhere.  It will be called @file{gtype-@var{lang}.h}, where
@var{lang} is the name of the subdirectory the language is contained in.

Plugins can add additional root tables.  Run the @code{gengtype}
utility in plugin mode as @code{gengtype -P pluginout.h @var{source-dir}
@var{file-list} @var{plugin*.c}} with your plugin files
@var{plugin*.c} using @code{GTY} to generate the @var{pluginout.h} file.
The GCC build tree is needed to be present in that mode.


@node Invoking the garbage collector
@section How to invoke the garbage collector
@cindex garbage collector, invocation
@findex ggc_collect

The GCC garbage collector GGC is only invoked explicitly. In contrast
with many other garbage collectors, it is not implicitly invoked by
allocation routines when a lot of memory has been consumed. So the
only way to have GGC reclaim storage it to call the @code{ggc_collect}
function explicitly.  This call is an expensive operation, as it may
have to scan the entire heap.  Beware that local variables (on the GCC
call stack) are not followed by such an invocation (as many other
garbage collectors do): you should reference all your data from static
or external @code{GTY}-ed variables, and it is advised to call
@code{ggc_collect} with a shallow call stack.  The GGC is an exact mark
and sweep garbage collector (so it does not scan the call stack for
pointers).  In practice GCC passes don't often call @code{ggc_collect}
themselves, because it is called by the pass manager between passes.

At the time of the @code{ggc_collect} call all pointers in the GC-marked
structures must be valid or @code{NULL}.  In practice this means that
there should not be uninitialized pointer fields in the structures even
if your code never reads or writes those fields at a particular
instance.  One way to ensure this is to use cleared versions of
allocators unless all the fields are initialized manually immediately
after allocation.

@node Troubleshooting
@section Troubleshooting the garbage collector
@cindex garbage collector, troubleshooting

With the current garbage collector implementation, most issues should
show up as GCC compilation errors.  Some of the most commonly
encountered issues are described below.

@itemize @bullet
@item Gengtype does not produce allocators for a @code{GTY}-marked type.
Gengtype checks if there is at least one possible path from GC roots to
at least one instance of each type before outputting allocators.  If
there is no such path, the @code{GTY} markers will be ignored and no
allocators will be output.  Solve this by making sure that there exists
at least one such path.  If creating it is unfeasible or raises a ``code
smell'', consider if you really must use GC for allocating such type.

@item Link-time errors about undefined @code{gt_ggc_r_foo_bar} and
similarly-named symbols.  Check if your @file{foo_bar} source file has
@code{#include "gt-foo_bar.h"} as its very last line.

@end itemize