Re: [PATCH v4] Add support for creating new types from the Python API

[Tom Tromey <tom@tromey.com>]

>>>>> "Matheus" == Matheus Branco Borella <dark.ryu.550@gmail.com> writes:

Matheus> The main drawback of using the `init_*_type` family over implementing type
Matheus> initialization by hand is that any type that's created gets immediately
Matheus> allocated on its owner's obstack, regardless of what its real lifetime
Matheus> requirements are. The main implication of this is that types that become
Matheus> unreachable will remain live for the lifetime of the owner.

Yeah.  gdb leaks a lot of types this way, actually.  We've collectively
put off implementing "type GC", though I do think there's a bug for it.

Matheus> +  ** Functions that allow creation of instances of gdb.Type, and a new
Matheus> +     class gdb.FloatFormat that may be used to create floating point
Matheus> +     types.  The functions that allow new type creation are:
Matheus> +      - gdb.init_type: Create a new type given a type code.
Matheus> +      - gdb.init_integer_type: Create a new integer type.
Matheus> +      - gdb.init_character_type: Create a new character type.
Matheus> +      - gdb.init_boolean_type: Create a new boolean type.
Matheus> +      - gdb.init_float_type: Create a new floating point type.
Matheus> +      - gdb.init_decfloat_type: Create a new decimal floating point type.
Matheus> +      - gdb.can_create_complex_type: Whether a type can be used to create a
Matheus> +          new complex type.
Matheus> +      - gdb.init_complex_type: Create a new complex type.
Matheus> +      - gdb.init_pointer_type: Create a new pointer type.
Matheus> +          * This allows creating pointers of arbitrary size.
Matheus> +      - gdb.init_fixed_point_type: Create a new fixed point type.

I don't really love the "init_" prefixes here.  Like, I get that these
are the names internally, but I don't think they really make sense
externally.

WDYT about "make_" instead?

Matheus> +@findex gdb.init_type
Matheus> +@defun gdb.init_type (owner, type_code, bit_size, name)
Matheus> +This function creates a new @code{gdb.Type} instance corresponding to a
Matheus> +type owned by the given @var{owner}, with the given @var{type_code},
Matheus> +@var{name} and size.
Matheus> +
Matheus> +@var{owner} must be a reference to either a @code{gdb.Objfile} or a
Matheus> +@code{gdb.Architecture} object.  These correspond to objfile and
Matheus> +architecture-owned types, respectively.
Matheus> +
Matheus> +@var{type_code} is one of the @code{TYPE_CODE_} constants defined in
Matheus> +@ref{Types In Python}.
Matheus> +
Matheus> +@var{bit_size} is the size of instances of the newly created type, in
Matheus> +bits. Currently, accepted values are limited to multiples of 8.
Matheus> +@end defun

Making any sort of type without filling in the details is probably a
recipe for crashes.

Is there a specific situation you needed this for?

Matheus> +@findex gdb.init_pointer_type
Matheus> +@defun gdb.init_pointer_type (owner, target, bit_size, name)
Matheus> +This function creates a new @code{gdb.Type} instance corresponding to a
Matheus> +pointer type that points to @var{target} and is owned by the given
Matheus> +@var{owner}, with the given @var{name} and size.
Matheus> +
Matheus> +@var{target} is a @code{gdb.Type} object, corresponding to the type
Matheus> +that will be pointed to by the newly created pointer type.
Matheus> +@end defun

I'm curious whether this one is really needed, because
gdb.Type.pointer() exists.

Like is there a case where you'd want a pointer type that doesn't match
the architecture somehow?  Seems weird and/or not useful.

Matheus> +/* Converts from a Python integer to a unsigned integer. */
Matheus> +
Matheus> +static bool
Matheus> +py_to_unsigned_int (PyObject *object, unsigned int *val)
Matheus> +{
Matheus> +  if (!PyObject_IsInstance (object, (PyObject*) &PyLong_Type))
Matheus> +    {
Matheus> +      PyErr_SetString (PyExc_TypeError, "value must be an integer");
Matheus> +      return false;
Matheus> +    }
Matheus> +
Matheus> +  long native_val = PyLong_AsLong (object);
Matheus> +  if (native_val > (long) UINT_MAX)
Matheus> +    {
Matheus> +      PyErr_SetString (PyExc_ValueError, "value is too large");
Matheus> +      return false;
Matheus> +    }
Matheus> +  if (native_val < 0)
Matheus> +    {
Matheus> +      PyErr_SetString (PyExc_ValueError,
Matheus> +		       "value must not be smaller than zero");
Matheus> +      return false;
Matheus> +    }
Matheus> +
Matheus> +  *val = (unsigned int) native_val;
Matheus> +  return true;

See gdb_py_int_as_long.
I think the type-check isn't really needed (probably) and some of the
other error-handling can be simplified.
There's also gdb_py_long_as_ulongest.

Matheus> +/* Functionality for creating new types accessible from python.
Matheus> +
Matheus> +   Copyright (C) 2008-2023 Free Software Foundation, Inc.

Forgot to mention this elsewhere but I think these dates are wrong.

Matheus> +/* An abstraction covering the objects types that can own a type object. */
Matheus> +
Matheus> +class type_storage_owner
Matheus> +{
Matheus> +public:
Matheus> +  /* Creates a new type owner from the given python object. If the object is
Matheus> +   * of a type that is not supported, the newly created instance will be
Matheus> +   * marked as invalid and nothing should be done with it. */
Matheus> +
Matheus> +  type_storage_owner (PyObject *owner)
Matheus> +  {
Matheus> +    if (gdbpy_is_architecture (owner))
Matheus> +      {
Matheus> +	this->kind = owner_kind::arch;
Matheus> +	this->owner.arch = arch_object_to_gdbarch (owner);
Matheus> +	return;
Matheus> +      }
Matheus> +
Matheus> +    this->kind = owner_kind::objfile;
Matheus> +    this->owner.objfile = objfile_object_to_objfile (owner);
Matheus> +    if (this->owner.objfile != nullptr)
Matheus> +	return;

It seems like this could all just create a type_allocator directly and
be simpler, like the 'kind' isn't needed.

Matheus> +
Matheus> +    this->kind = owner_kind::none;
Matheus> +    PyErr_SetString(PyExc_TypeError, "unsupported owner type");

Spaces before parens in a lot of spots...

Matheus> +    /* Should never be reached, but it's better to fail in a safe way than try
Matheus> +     * to instance the allocator with arbitraty parameters here. */
Matheus> +    abort ();

gdb uses gdb_assert_not_reached instead.

Matheus> +  /* Get a reference to the owner's obstack. */
Matheus> +
Matheus> +  obstack *get_obstack ()
Matheus> +  {

I think the uses of this could probably use TYPE_ALLOC instead.

Matheus> +  struct gdbarch *get_arch ()
Matheus> +  {

This could use the type allocator's arch.

Matheus> +  enum class owner_kind { arch, objfile, none };

Is the none case really possible?
It might be better to just throw an exception from the constructor or
during argument validation or something like that.

thanks,
Tom