2021-11-21 23:59:44 +00:00
|
|
|
// Copyright (c) Meta Platforms, Inc. and affiliates.
|
2022-11-02 00:05:16 +00:00
|
|
|
// SPDX-License-Identifier: LGPL-2.1-or-later
|
2020-05-15 23:13:02 +01:00
|
|
|
|
2021-02-23 22:31:01 +00:00
|
|
|
#include <limits.h>
|
2020-09-24 00:02:02 +01:00
|
|
|
#include <stdlib.h>
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
#include <string.h>
|
|
|
|
|
2021-08-24 01:07:26 +01:00
|
|
|
#include "array.h"
|
2020-09-24 00:02:02 +01:00
|
|
|
#include "error.h"
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
#include "hash_table.h"
|
|
|
|
#include "language.h"
|
2020-12-18 19:01:29 +00:00
|
|
|
#include "lazy_object.h"
|
2020-04-23 00:23:26 +01:00
|
|
|
#include "program.h"
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
#include "type.h"
|
2020-09-24 00:02:02 +01:00
|
|
|
#include "util.h"
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
|
2019-04-28 16:42:28 +01:00
|
|
|
const char * const drgn_type_kind_spelling[] = {
|
|
|
|
[DRGN_TYPE_VOID] = "void",
|
|
|
|
[DRGN_TYPE_INT] = "int",
|
|
|
|
[DRGN_TYPE_BOOL] = "bool",
|
|
|
|
[DRGN_TYPE_FLOAT] = "float",
|
|
|
|
[DRGN_TYPE_STRUCT] = "struct",
|
|
|
|
[DRGN_TYPE_UNION] = "union",
|
2019-11-15 01:12:47 +00:00
|
|
|
[DRGN_TYPE_CLASS] = "class",
|
2019-04-28 16:42:28 +01:00
|
|
|
[DRGN_TYPE_ENUM] = "enum",
|
|
|
|
[DRGN_TYPE_TYPEDEF] = "typedef",
|
|
|
|
[DRGN_TYPE_POINTER] = "pointer",
|
|
|
|
[DRGN_TYPE_ARRAY] = "array",
|
|
|
|
[DRGN_TYPE_FUNCTION] = "function",
|
|
|
|
};
|
|
|
|
|
2020-04-23 00:23:26 +01:00
|
|
|
/**
|
|
|
|
* Names of primitive types.
|
|
|
|
*
|
|
|
|
* In some languages, like C, the same primitive type can be spelled in multiple
|
|
|
|
* ways. For example, "int" can also be spelled "signed int" or "int signed".
|
|
|
|
*
|
|
|
|
* This maps each @ref drgn_primitive_type to a ``NULL``-terminated array of the
|
|
|
|
* different ways to spell that type. The spelling at index zero is the
|
|
|
|
* preferred spelling.
|
|
|
|
*/
|
|
|
|
static const char * const * const
|
2019-04-20 00:05:19 +01:00
|
|
|
drgn_primitive_type_spellings[DRGN_PRIMITIVE_TYPE_NUM] = {
|
|
|
|
[DRGN_C_TYPE_VOID] = (const char * []){ "void", NULL, },
|
|
|
|
[DRGN_C_TYPE_CHAR] = (const char * []){ "char", NULL, },
|
|
|
|
[DRGN_C_TYPE_SIGNED_CHAR] = (const char * []){
|
|
|
|
"signed char", "char signed", NULL,
|
|
|
|
},
|
|
|
|
[DRGN_C_TYPE_UNSIGNED_CHAR] = (const char * []){
|
|
|
|
"unsigned char", "char unsigned", NULL,
|
|
|
|
},
|
|
|
|
[DRGN_C_TYPE_SHORT] = (const char * []){
|
|
|
|
"short", "signed short", "short signed", "short int",
|
|
|
|
"int short", "signed short int", "signed int short",
|
|
|
|
"short signed int", "short int signed", "int signed short",
|
|
|
|
"int short signed", NULL,
|
|
|
|
},
|
|
|
|
[DRGN_C_TYPE_UNSIGNED_SHORT] = (const char * []){
|
|
|
|
"unsigned short", "short unsigned", "unsigned short int",
|
|
|
|
"unsigned int short", "short unsigned int",
|
|
|
|
"short int unsigned", "int unsigned short",
|
|
|
|
"int short unsigned", NULL,
|
|
|
|
},
|
|
|
|
[DRGN_C_TYPE_INT] = (const char * []){
|
|
|
|
"int", "signed", "signed int", "int signed", NULL,
|
|
|
|
},
|
|
|
|
[DRGN_C_TYPE_UNSIGNED_INT] = (const char * []){
|
|
|
|
"unsigned int", "int unsigned", "unsigned", NULL,
|
|
|
|
},
|
|
|
|
[DRGN_C_TYPE_LONG] = (const char * []){
|
|
|
|
"long", "signed long", "long signed", "long int", "int long",
|
|
|
|
"signed long int", "signed int long", "long signed int",
|
|
|
|
"long int signed", "int signed long", "int long signed", NULL,
|
|
|
|
},
|
|
|
|
[DRGN_C_TYPE_UNSIGNED_LONG] = (const char * []){
|
|
|
|
"unsigned long", "long unsigned", "unsigned long int",
|
|
|
|
"unsigned int long", "long unsigned int", "long int unsigned",
|
|
|
|
"int unsigned long", "int long unsigned", NULL,
|
|
|
|
},
|
|
|
|
[DRGN_C_TYPE_LONG_LONG] = (const char * []){
|
|
|
|
"long long", "signed long long", "long signed long",
|
|
|
|
"long long signed", "long long int", "long int long",
|
|
|
|
"int long long", "signed long long int", "signed long int long",
|
|
|
|
"signed int long long", "long signed long int",
|
|
|
|
"long signed int long", "long long signed int",
|
|
|
|
"long long int signed", "long int signed long",
|
|
|
|
"long int long signed", "int signed long long",
|
|
|
|
"int long signed long", "int long long signed", NULL,
|
|
|
|
},
|
|
|
|
[DRGN_C_TYPE_UNSIGNED_LONG_LONG] = (const char * []){
|
|
|
|
"unsigned long long", "long unsigned long",
|
|
|
|
"long long unsigned", "unsigned long long int",
|
|
|
|
"unsigned long int long", "unsigned int long long",
|
|
|
|
"long unsigned long int", "long unsigned int long",
|
|
|
|
"long long unsigned int", "long long int unsigned",
|
|
|
|
"long int unsigned long", "long int long unsigned",
|
|
|
|
"int unsigned long long", "int long unsigned long",
|
|
|
|
"int long long unsigned", NULL,
|
|
|
|
},
|
|
|
|
[DRGN_C_TYPE_BOOL] = (const char * []){ "_Bool", NULL, },
|
|
|
|
[DRGN_C_TYPE_FLOAT] = (const char * []){ "float", NULL, },
|
|
|
|
[DRGN_C_TYPE_DOUBLE] = (const char * []){ "double", NULL, },
|
|
|
|
[DRGN_C_TYPE_LONG_DOUBLE] = (const char * []){
|
|
|
|
"long double", "double long", NULL,
|
|
|
|
},
|
|
|
|
[DRGN_C_TYPE_SIZE_T] = (const char * []){ "size_t", NULL, },
|
|
|
|
[DRGN_C_TYPE_PTRDIFF_T] = (const char * []){ "ptrdiff_t", NULL, },
|
|
|
|
};
|
|
|
|
|
2020-04-23 00:23:26 +01:00
|
|
|
/**
|
|
|
|
* Mapping from a @ref drgn_type_primitive to the corresponding @ref
|
|
|
|
* drgn_type_kind.
|
|
|
|
*/
|
|
|
|
static const enum drgn_type_kind
|
2019-04-20 00:05:19 +01:00
|
|
|
drgn_primitive_type_kind[DRGN_PRIMITIVE_TYPE_NUM + 1] = {
|
|
|
|
[DRGN_C_TYPE_CHAR] = DRGN_TYPE_INT,
|
|
|
|
[DRGN_C_TYPE_SIGNED_CHAR] = DRGN_TYPE_INT,
|
|
|
|
[DRGN_C_TYPE_UNSIGNED_CHAR] = DRGN_TYPE_INT,
|
|
|
|
[DRGN_C_TYPE_SHORT] = DRGN_TYPE_INT,
|
|
|
|
[DRGN_C_TYPE_UNSIGNED_SHORT] = DRGN_TYPE_INT,
|
|
|
|
[DRGN_C_TYPE_INT] = DRGN_TYPE_INT,
|
|
|
|
[DRGN_C_TYPE_UNSIGNED_INT] = DRGN_TYPE_INT,
|
|
|
|
[DRGN_C_TYPE_LONG] = DRGN_TYPE_INT,
|
|
|
|
[DRGN_C_TYPE_UNSIGNED_LONG] = DRGN_TYPE_INT,
|
|
|
|
[DRGN_C_TYPE_LONG_LONG] = DRGN_TYPE_INT,
|
|
|
|
[DRGN_C_TYPE_UNSIGNED_LONG_LONG] = DRGN_TYPE_INT,
|
|
|
|
[DRGN_C_TYPE_BOOL] = DRGN_TYPE_BOOL,
|
|
|
|
[DRGN_C_TYPE_FLOAT] = DRGN_TYPE_FLOAT,
|
|
|
|
[DRGN_C_TYPE_DOUBLE] = DRGN_TYPE_FLOAT,
|
|
|
|
[DRGN_C_TYPE_LONG_DOUBLE] = DRGN_TYPE_FLOAT,
|
|
|
|
[DRGN_C_TYPE_SIZE_T] = DRGN_TYPE_TYPEDEF,
|
|
|
|
[DRGN_C_TYPE_PTRDIFF_T] = DRGN_TYPE_TYPEDEF,
|
|
|
|
[DRGN_C_TYPE_VOID] = DRGN_TYPE_VOID,
|
|
|
|
[DRGN_NOT_PRIMITIVE_TYPE] = -1,
|
|
|
|
};
|
|
|
|
|
|
|
|
/** Return whether a primitive type is always a signed integer type. */
|
|
|
|
static inline bool
|
|
|
|
drgn_primitive_type_is_signed(enum drgn_primitive_type primitive)
|
|
|
|
{
|
|
|
|
switch (primitive) {
|
|
|
|
case DRGN_C_TYPE_SIGNED_CHAR:
|
|
|
|
case DRGN_C_TYPE_SHORT:
|
|
|
|
case DRGN_C_TYPE_INT:
|
|
|
|
case DRGN_C_TYPE_LONG:
|
|
|
|
case DRGN_C_TYPE_LONG_LONG:
|
|
|
|
case DRGN_C_TYPE_PTRDIFF_T:
|
|
|
|
return true;
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2020-10-13 00:06:23 +01:00
|
|
|
static struct hash_pair
|
|
|
|
drgn_member_key_hash_pair(const struct drgn_member_key *key)
|
2020-04-23 00:23:26 +01:00
|
|
|
{
|
|
|
|
size_t hash;
|
|
|
|
if (key->name)
|
2020-10-13 00:06:23 +01:00
|
|
|
hash = hash_bytes(key->name, key->name_len);
|
2020-04-23 00:23:26 +01:00
|
|
|
else
|
|
|
|
hash = 0;
|
|
|
|
hash = hash_combine((uintptr_t)key->type, hash);
|
|
|
|
return hash_pair_from_avalanching_hash(hash);
|
|
|
|
}
|
|
|
|
|
2020-10-13 00:06:23 +01:00
|
|
|
static bool drgn_member_key_eq(const struct drgn_member_key *a,
|
|
|
|
const struct drgn_member_key *b)
|
2020-04-23 00:23:26 +01:00
|
|
|
{
|
|
|
|
return (a->type == b->type && a->name_len == b->name_len &&
|
|
|
|
(!a->name_len || memcmp(a->name, b->name, a->name_len) == 0));
|
|
|
|
}
|
|
|
|
|
2021-10-11 21:01:07 +01:00
|
|
|
DEFINE_HASH_MAP_FUNCTIONS(drgn_member_map, drgn_member_key_hash_pair,
|
|
|
|
drgn_member_key_eq)
|
2020-04-23 00:23:26 +01:00
|
|
|
|
2021-10-11 21:01:07 +01:00
|
|
|
DEFINE_HASH_SET_FUNCTIONS(drgn_type_set, ptr_key_hash_pair, scalar_key_eq)
|
2020-04-23 00:23:26 +01:00
|
|
|
|
2020-12-18 19:01:29 +00:00
|
|
|
LIBDRGN_PUBLIC struct drgn_error *
|
|
|
|
drgn_member_object(struct drgn_type_member *member,
|
|
|
|
const struct drgn_object **ret)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
2020-12-18 19:01:29 +00:00
|
|
|
struct drgn_error *err = drgn_lazy_object_evaluate(&member->object);
|
|
|
|
if (!err)
|
|
|
|
*ret = &member->object.obj;
|
|
|
|
return err;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
2020-12-18 19:01:29 +00:00
|
|
|
LIBDRGN_PUBLIC struct drgn_error *
|
|
|
|
drgn_member_type(struct drgn_type_member *member,
|
|
|
|
struct drgn_qualified_type *type_ret,
|
|
|
|
uint64_t *bit_field_size_ret)
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
{
|
2020-12-18 19:01:29 +00:00
|
|
|
struct drgn_error *err = drgn_lazy_object_evaluate(&member->object);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
*type_ret = drgn_object_qualified_type(&member->object.obj);
|
|
|
|
if (bit_field_size_ret) {
|
|
|
|
if (member->object.obj.is_bit_field)
|
|
|
|
*bit_field_size_ret = member->object.obj.bit_size;
|
|
|
|
else
|
|
|
|
*bit_field_size_ret = 0;
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
}
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
LIBDRGN_PUBLIC struct drgn_error *
|
2020-12-18 19:01:29 +00:00
|
|
|
drgn_parameter_default_argument(struct drgn_type_parameter *parameter,
|
|
|
|
const struct drgn_object **ret)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
2020-12-18 19:01:29 +00:00
|
|
|
struct drgn_error *err =
|
|
|
|
drgn_lazy_object_evaluate(¶meter->default_argument);
|
|
|
|
if (!err)
|
|
|
|
*ret = ¶meter->default_argument.obj;
|
|
|
|
return err;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
LIBDRGN_PUBLIC struct drgn_error *
|
|
|
|
drgn_parameter_type(struct drgn_type_parameter *parameter,
|
|
|
|
struct drgn_qualified_type *ret)
|
|
|
|
{
|
2020-12-18 19:01:29 +00:00
|
|
|
struct drgn_error *err =
|
|
|
|
drgn_lazy_object_evaluate(¶meter->default_argument);
|
|
|
|
if (!err)
|
|
|
|
*ret = drgn_object_qualified_type(¶meter->default_argument.obj);
|
|
|
|
return err;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
2021-01-09 01:28:27 +00:00
|
|
|
LIBDRGN_PUBLIC struct drgn_error *
|
|
|
|
drgn_template_parameter_type(struct drgn_type_template_parameter *parameter,
|
|
|
|
struct drgn_qualified_type *ret)
|
|
|
|
{
|
|
|
|
struct drgn_error *err =
|
|
|
|
drgn_lazy_object_evaluate(¶meter->argument);
|
|
|
|
if (!err)
|
|
|
|
*ret = drgn_object_qualified_type(¶meter->argument.obj);
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
|
|
|
LIBDRGN_PUBLIC struct drgn_error *
|
|
|
|
drgn_template_parameter_object(struct drgn_type_template_parameter *parameter,
|
|
|
|
const struct drgn_object **ret)
|
|
|
|
{
|
|
|
|
struct drgn_error *err =
|
|
|
|
drgn_lazy_object_evaluate(¶meter->argument);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
if (parameter->argument.obj.kind == DRGN_OBJECT_ABSENT)
|
|
|
|
*ret = NULL;
|
|
|
|
else
|
|
|
|
*ret = ¶meter->argument.obj;
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2020-10-13 00:06:23 +01:00
|
|
|
static struct hash_pair
|
|
|
|
drgn_type_dedupe_hash_pair(struct drgn_type * const *entry)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_type *type = *entry;
|
|
|
|
size_t hash = hash_combine(drgn_type_kind(type),
|
2021-01-02 09:41:24 +00:00
|
|
|
drgn_type_is_complete(type));
|
|
|
|
hash = hash_combine(hash, (uintptr_t)drgn_type_language(type));
|
2020-10-13 00:06:23 +01:00
|
|
|
if (drgn_type_has_name(type))
|
|
|
|
hash = hash_combine(hash, hash_c_string(drgn_type_name(type)));
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
if (drgn_type_has_size(type))
|
|
|
|
hash = hash_combine(hash, drgn_type_size(type));
|
|
|
|
if (drgn_type_has_is_signed(type))
|
|
|
|
hash = hash_combine(hash, drgn_type_is_signed(type));
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
if (drgn_type_has_little_endian(type))
|
|
|
|
hash = hash_combine(hash, drgn_type_little_endian(type));
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
const char *tag;
|
|
|
|
if (drgn_type_has_tag(type) && (tag = drgn_type_tag(type)))
|
2020-10-13 00:06:23 +01:00
|
|
|
hash = hash_combine(hash, hash_c_string(tag));
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
if (drgn_type_has_type(type)) {
|
|
|
|
struct drgn_qualified_type qualified_type =
|
|
|
|
drgn_type_type(type);
|
|
|
|
hash = hash_combine(hash, (uintptr_t)qualified_type.type);
|
|
|
|
hash = hash_combine(hash, qualified_type.qualifiers);
|
|
|
|
}
|
2021-01-02 09:41:24 +00:00
|
|
|
if (drgn_type_has_length(type))
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
hash = hash_combine(hash, drgn_type_length(type));
|
2021-01-02 09:41:24 +00:00
|
|
|
if (drgn_type_has_is_variadic(type))
|
|
|
|
hash = hash_combine(hash, drgn_type_is_variadic(type));
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
return hash_pair_from_avalanching_hash(hash);
|
|
|
|
}
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
static bool drgn_type_dedupe_eq(struct drgn_type * const *entry_a,
|
|
|
|
struct drgn_type * const *entry_b)
|
|
|
|
{
|
|
|
|
struct drgn_type *a = *entry_a;
|
|
|
|
struct drgn_type *b = *entry_b;
|
|
|
|
|
|
|
|
if (drgn_type_kind(a) != drgn_type_kind(b) ||
|
2021-01-02 09:41:24 +00:00
|
|
|
drgn_type_is_complete(a) != drgn_type_is_complete(b) ||
|
|
|
|
drgn_type_language(a) != drgn_type_language(b))
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
return false;
|
|
|
|
if (drgn_type_has_name(a) &&
|
|
|
|
strcmp(drgn_type_name(a), drgn_type_name(b)) != 0)
|
|
|
|
return false;
|
|
|
|
if (drgn_type_has_size(a) && drgn_type_size(a) != drgn_type_size(b))
|
|
|
|
return false;
|
|
|
|
if (drgn_type_has_is_signed(a) &&
|
|
|
|
drgn_type_is_signed(a) != drgn_type_is_signed(b))
|
|
|
|
return false;
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
if (drgn_type_has_little_endian(a) &&
|
|
|
|
drgn_type_little_endian(a) != drgn_type_little_endian(b))
|
|
|
|
return false;
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
if (drgn_type_has_tag(a)) {
|
|
|
|
const char *tag_a = drgn_type_tag(a);
|
|
|
|
const char *tag_b = drgn_type_tag(b);
|
|
|
|
if ((!tag_a != !tag_b) || (tag_a && strcmp(tag_a, tag_b) != 0))
|
|
|
|
return false;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
if (drgn_type_has_type(a)) {
|
|
|
|
struct drgn_qualified_type type_a = drgn_type_type(a);
|
|
|
|
struct drgn_qualified_type type_b = drgn_type_type(b);
|
|
|
|
if (type_a.type != type_b.type ||
|
|
|
|
type_a.qualifiers != type_b.qualifiers)
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
if (drgn_type_has_length(a) &&
|
|
|
|
drgn_type_length(a) != drgn_type_length(b))
|
|
|
|
return false;
|
2021-01-02 09:41:24 +00:00
|
|
|
if (drgn_type_has_is_variadic(a) &&
|
|
|
|
drgn_type_is_variadic(a) != drgn_type_is_variadic(b))
|
|
|
|
return false;
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
return true;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
/*
|
2021-01-09 01:28:27 +00:00
|
|
|
* We don't deduplicate types with members, parameters, template parameters, or
|
|
|
|
* enumerators, so the hash and comparison functions ignore those.
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
*/
|
2021-10-11 21:01:07 +01:00
|
|
|
DEFINE_HASH_SET_FUNCTIONS(drgn_dedupe_type_set, drgn_type_dedupe_hash_pair,
|
|
|
|
drgn_type_dedupe_eq)
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
|
|
|
|
DEFINE_VECTOR_FUNCTIONS(drgn_typep_vector)
|
|
|
|
|
|
|
|
static struct drgn_error *find_or_create_type(struct drgn_type *key,
|
|
|
|
struct drgn_type **ret)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_program *prog = key->_private.program;
|
|
|
|
struct hash_pair hp = drgn_dedupe_type_set_hash(&key);
|
|
|
|
struct drgn_dedupe_type_set_iterator it =
|
|
|
|
drgn_dedupe_type_set_search_hashed(&prog->dedupe_types, &key,
|
|
|
|
hp);
|
|
|
|
if (it.entry) {
|
|
|
|
*ret = *it.entry;
|
|
|
|
return NULL;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
|
|
|
|
struct drgn_type *type = malloc(sizeof(*type));
|
|
|
|
if (!type)
|
|
|
|
return &drgn_enomem;
|
|
|
|
|
|
|
|
*type = *key;
|
2021-11-17 09:18:49 +00:00
|
|
|
if (drgn_dedupe_type_set_insert_searched(&prog->dedupe_types, &type, hp,
|
|
|
|
NULL) < 0) {
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
free(type);
|
|
|
|
return &drgn_enomem;
|
|
|
|
}
|
|
|
|
*ret = type;
|
|
|
|
return NULL;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_type *drgn_void_type(struct drgn_program *prog,
|
|
|
|
const struct drgn_language *lang)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
if (!lang)
|
|
|
|
lang = drgn_program_language(prog);
|
2022-02-16 20:43:36 +00:00
|
|
|
return &prog->void_types[lang->number];
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
}
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
static struct drgn_error *
|
|
|
|
drgn_byte_order_to_little_endian(struct drgn_program *prog,
|
|
|
|
enum drgn_byte_order byte_order, bool *ret)
|
|
|
|
{
|
|
|
|
SWITCH_ENUM_DEFAULT(byte_order,
|
|
|
|
case DRGN_BIG_ENDIAN:
|
|
|
|
*ret = false;
|
|
|
|
return NULL;
|
|
|
|
case DRGN_LITTLE_ENDIAN:
|
|
|
|
*ret = true;
|
|
|
|
return NULL;
|
|
|
|
case DRGN_PROGRAM_ENDIAN:
|
|
|
|
return drgn_program_is_little_endian(prog, ret);
|
|
|
|
default:
|
|
|
|
return drgn_error_create(DRGN_ERROR_INVALID_ARGUMENT,
|
|
|
|
"invalid byte order");
|
|
|
|
)
|
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_error *drgn_int_type_create(struct drgn_program *prog,
|
|
|
|
const char *name, uint64_t size,
|
|
|
|
bool is_signed,
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
enum drgn_byte_order byte_order,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
const struct drgn_language *lang,
|
|
|
|
struct drgn_type **ret)
|
|
|
|
{
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
struct drgn_error *err;
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
enum drgn_primitive_type primitive = c_parse_specifier_list(name);
|
|
|
|
if (drgn_primitive_type_kind[primitive] == DRGN_TYPE_INT &&
|
|
|
|
(primitive == DRGN_C_TYPE_CHAR ||
|
|
|
|
is_signed == drgn_primitive_type_is_signed(primitive)))
|
|
|
|
name = drgn_primitive_type_spellings[primitive][0];
|
|
|
|
else
|
|
|
|
primitive = DRGN_NOT_PRIMITIVE_TYPE;
|
|
|
|
|
|
|
|
struct drgn_type key = {
|
|
|
|
{
|
|
|
|
.kind = DRGN_TYPE_INT,
|
|
|
|
.is_complete = true,
|
|
|
|
.primitive = primitive,
|
|
|
|
.name = name,
|
|
|
|
.size = size,
|
|
|
|
.is_signed = is_signed,
|
|
|
|
.program = prog,
|
|
|
|
.language = lang ? lang : drgn_program_language(prog),
|
|
|
|
}
|
|
|
|
};
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
err = drgn_byte_order_to_little_endian(prog, byte_order,
|
|
|
|
&key._private.little_endian);
|
|
|
|
if (err)
|
|
|
|
return err;
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
return find_or_create_type(&key, ret);
|
|
|
|
}
|
|
|
|
|
|
|
|
struct drgn_error *drgn_bool_type_create(struct drgn_program *prog,
|
|
|
|
const char *name, uint64_t size,
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
enum drgn_byte_order byte_order,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
const struct drgn_language *lang,
|
|
|
|
struct drgn_type **ret)
|
|
|
|
{
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
struct drgn_error *err;
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
enum drgn_primitive_type primitive = c_parse_specifier_list(name);
|
|
|
|
if (primitive == DRGN_C_TYPE_BOOL)
|
|
|
|
name = drgn_primitive_type_spellings[DRGN_C_TYPE_BOOL][0];
|
|
|
|
else
|
|
|
|
primitive = DRGN_NOT_PRIMITIVE_TYPE;
|
|
|
|
|
|
|
|
struct drgn_type key = {
|
|
|
|
{
|
|
|
|
.kind = DRGN_TYPE_BOOL,
|
|
|
|
.is_complete = true,
|
|
|
|
.primitive = primitive,
|
|
|
|
.name = name,
|
|
|
|
.size = size,
|
|
|
|
.program = prog,
|
|
|
|
.language = lang ? lang : drgn_program_language(prog),
|
|
|
|
}
|
|
|
|
};
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
err = drgn_byte_order_to_little_endian(prog, byte_order,
|
|
|
|
&key._private.little_endian);
|
|
|
|
if (err)
|
|
|
|
return err;
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
return find_or_create_type(&key, ret);
|
|
|
|
}
|
|
|
|
|
|
|
|
struct drgn_error *drgn_float_type_create(struct drgn_program *prog,
|
|
|
|
const char *name, uint64_t size,
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
enum drgn_byte_order byte_order,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
const struct drgn_language *lang,
|
|
|
|
struct drgn_type **ret)
|
|
|
|
{
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
struct drgn_error *err;
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
enum drgn_primitive_type primitive = c_parse_specifier_list(name);
|
|
|
|
if (drgn_primitive_type_kind[primitive] == DRGN_TYPE_FLOAT)
|
|
|
|
name = drgn_primitive_type_spellings[primitive][0];
|
|
|
|
else
|
|
|
|
primitive = DRGN_NOT_PRIMITIVE_TYPE;
|
|
|
|
|
|
|
|
struct drgn_type key = {
|
|
|
|
{
|
|
|
|
.kind = DRGN_TYPE_FLOAT,
|
|
|
|
.is_complete = true,
|
|
|
|
.primitive = primitive,
|
|
|
|
.name = name,
|
|
|
|
.size = size,
|
|
|
|
.program = prog,
|
|
|
|
.language = lang ? lang : drgn_program_language(prog),
|
|
|
|
}
|
|
|
|
};
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
err = drgn_byte_order_to_little_endian(prog, byte_order,
|
|
|
|
&key._private.little_endian);
|
|
|
|
if (err)
|
|
|
|
return err;
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
return find_or_create_type(&key, ret);
|
|
|
|
}
|
|
|
|
|
2021-01-09 01:28:27 +00:00
|
|
|
DEFINE_VECTOR_FUNCTIONS(drgn_type_template_parameter_vector)
|
|
|
|
|
|
|
|
static void
|
|
|
|
drgn_template_parameters_builder_init(struct drgn_template_parameters_builder *builder,
|
|
|
|
struct drgn_program *prog)
|
|
|
|
{
|
|
|
|
builder->prog = prog;
|
|
|
|
drgn_type_template_parameter_vector_init(&builder->parameters);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void
|
|
|
|
drgn_template_parameters_builder_deinit(struct drgn_template_parameters_builder *builder)
|
|
|
|
{
|
|
|
|
for (size_t i = 0; i < builder->parameters.size; i++)
|
|
|
|
drgn_lazy_object_deinit(&builder->parameters.data[i].argument);
|
|
|
|
drgn_type_template_parameter_vector_deinit(&builder->parameters);
|
|
|
|
}
|
|
|
|
|
|
|
|
struct drgn_error *
|
|
|
|
drgn_template_parameters_builder_add(struct drgn_template_parameters_builder *builder,
|
|
|
|
const union drgn_lazy_object *argument,
|
|
|
|
const char *name, bool is_default)
|
|
|
|
{
|
|
|
|
struct drgn_error *err = drgn_lazy_object_check_prog(argument,
|
|
|
|
builder->prog);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
struct drgn_type_template_parameter *parameter =
|
|
|
|
drgn_type_template_parameter_vector_append_entry(&builder->parameters);
|
|
|
|
if (!parameter)
|
|
|
|
return &drgn_enomem;
|
|
|
|
parameter->argument = *argument;
|
|
|
|
parameter->name = name;
|
|
|
|
parameter->is_default = is_default;
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
DEFINE_VECTOR_FUNCTIONS(drgn_type_member_vector)
|
|
|
|
|
|
|
|
void drgn_compound_type_builder_init(struct drgn_compound_type_builder *builder,
|
|
|
|
struct drgn_program *prog,
|
|
|
|
enum drgn_type_kind kind)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
assert(kind == DRGN_TYPE_STRUCT ||
|
|
|
|
kind == DRGN_TYPE_UNION ||
|
|
|
|
kind == DRGN_TYPE_CLASS);
|
2021-01-09 01:28:27 +00:00
|
|
|
drgn_template_parameters_builder_init(&builder->template_builder, prog);
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
builder->kind = kind;
|
|
|
|
drgn_type_member_vector_init(&builder->members);
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
void
|
|
|
|
drgn_compound_type_builder_deinit(struct drgn_compound_type_builder *builder)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
for (size_t i = 0; i < builder->members.size; i++)
|
2020-12-18 19:01:29 +00:00
|
|
|
drgn_lazy_object_deinit(&builder->members.data[i].object);
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
drgn_type_member_vector_deinit(&builder->members);
|
2021-01-09 01:28:27 +00:00
|
|
|
drgn_template_parameters_builder_deinit(&builder->template_builder);
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_error *
|
|
|
|
drgn_compound_type_builder_add_member(struct drgn_compound_type_builder *builder,
|
2020-12-18 19:01:29 +00:00
|
|
|
const union drgn_lazy_object *object,
|
|
|
|
const char *name, uint64_t bit_offset)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
2021-01-09 01:28:27 +00:00
|
|
|
struct drgn_error *err =
|
|
|
|
drgn_lazy_object_check_prog(object,
|
|
|
|
builder->template_builder.prog);
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
struct drgn_type_member *member =
|
|
|
|
drgn_type_member_vector_append_entry(&builder->members);
|
|
|
|
if (!member)
|
|
|
|
return &drgn_enomem;
|
2020-12-18 19:01:29 +00:00
|
|
|
member->object = *object;
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
member->name = name;
|
|
|
|
member->bit_offset = bit_offset;
|
|
|
|
return NULL;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_error *
|
|
|
|
drgn_compound_type_create(struct drgn_compound_type_builder *builder,
|
2021-01-09 01:28:27 +00:00
|
|
|
const char *tag, uint64_t size, bool is_complete,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
const struct drgn_language *lang,
|
|
|
|
struct drgn_type **ret)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
2021-01-07 01:13:56 +00:00
|
|
|
struct drgn_error *err;
|
2021-01-09 01:28:27 +00:00
|
|
|
struct drgn_program *prog = builder->template_builder.prog;
|
2021-01-07 01:13:56 +00:00
|
|
|
|
2021-01-09 01:28:27 +00:00
|
|
|
if (!is_complete) {
|
|
|
|
if (builder->members.size) {
|
|
|
|
return drgn_error_create(DRGN_ERROR_INVALID_ARGUMENT,
|
|
|
|
"incomplete type must not have members");
|
|
|
|
}
|
|
|
|
if (size != 0) {
|
|
|
|
return drgn_error_create(DRGN_ERROR_INVALID_ARGUMENT,
|
|
|
|
"size of incomplete type must be zero");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!builder->members.size &&
|
|
|
|
!builder->template_builder.parameters.size) {
|
2021-01-02 09:41:24 +00:00
|
|
|
struct drgn_type key = {
|
|
|
|
{
|
|
|
|
.kind = builder->kind,
|
2021-01-09 01:28:27 +00:00
|
|
|
.is_complete = is_complete,
|
2021-01-02 09:41:24 +00:00
|
|
|
.primitive = DRGN_NOT_PRIMITIVE_TYPE,
|
|
|
|
.tag = tag,
|
|
|
|
.size = size,
|
2021-01-09 01:28:27 +00:00
|
|
|
.program = prog,
|
2021-01-02 09:41:24 +00:00
|
|
|
.language =
|
2021-01-09 01:28:27 +00:00
|
|
|
lang ? lang : drgn_program_language(prog),
|
2021-01-02 09:41:24 +00:00
|
|
|
}
|
|
|
|
};
|
2021-01-07 01:13:56 +00:00
|
|
|
err = find_or_create_type(&key, ret);
|
|
|
|
if (!err)
|
|
|
|
drgn_type_member_vector_deinit(&builder->members);
|
|
|
|
return err;
|
2021-01-02 09:41:24 +00:00
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_type *type = malloc(sizeof(*type));
|
|
|
|
if (!type)
|
|
|
|
return &drgn_enomem;
|
2021-01-09 01:28:27 +00:00
|
|
|
if (!drgn_typep_vector_append(&prog->created_types, &type)) {
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
free(type);
|
|
|
|
return &drgn_enomem;
|
|
|
|
}
|
|
|
|
|
|
|
|
drgn_type_member_vector_shrink_to_fit(&builder->members);
|
2021-01-09 01:28:27 +00:00
|
|
|
drgn_type_template_parameter_vector_shrink_to_fit(&builder->template_builder.parameters);
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
|
|
|
|
type->_private.kind = builder->kind;
|
2021-01-09 01:28:27 +00:00
|
|
|
type->_private.is_complete = is_complete;
|
2019-04-20 00:05:19 +01:00
|
|
|
type->_private.primitive = DRGN_NOT_PRIMITIVE_TYPE;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
type->_private.tag = tag;
|
|
|
|
type->_private.size = size;
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
type->_private.members = builder->members.data;
|
|
|
|
type->_private.num_members = builder->members.size;
|
2021-01-09 01:28:27 +00:00
|
|
|
type->_private.template_parameters =
|
|
|
|
builder->template_builder.parameters.data;
|
|
|
|
type->_private.num_template_parameters =
|
|
|
|
builder->template_builder.parameters.size;
|
|
|
|
type->_private.program = prog;
|
|
|
|
type->_private.language = lang ? lang : drgn_program_language(prog);
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
*ret = type;
|
|
|
|
return NULL;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
DEFINE_VECTOR_FUNCTIONS(drgn_type_enumerator_vector)
|
|
|
|
|
|
|
|
void drgn_enum_type_builder_init(struct drgn_enum_type_builder *builder,
|
|
|
|
struct drgn_program *prog)
|
2019-11-15 01:12:47 +00:00
|
|
|
{
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
builder->prog = prog;
|
|
|
|
drgn_type_enumerator_vector_init(&builder->enumerators);
|
2019-11-15 01:12:47 +00:00
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
void drgn_enum_type_builder_deinit(struct drgn_enum_type_builder *builder)
|
2019-11-15 01:12:47 +00:00
|
|
|
{
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
drgn_type_enumerator_vector_deinit(&builder->enumerators);
|
|
|
|
}
|
|
|
|
|
|
|
|
struct drgn_error *
|
|
|
|
drgn_enum_type_builder_add_signed(struct drgn_enum_type_builder *builder,
|
|
|
|
const char *name, int64_t svalue)
|
|
|
|
{
|
|
|
|
struct drgn_type_enumerator *enumerator =
|
|
|
|
drgn_type_enumerator_vector_append_entry(&builder->enumerators);
|
|
|
|
if (!enumerator)
|
|
|
|
return &drgn_enomem;
|
|
|
|
enumerator->name = name;
|
|
|
|
enumerator->svalue = svalue;
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
struct drgn_error *
|
|
|
|
drgn_enum_type_builder_add_unsigned(struct drgn_enum_type_builder *builder,
|
|
|
|
const char *name, uint64_t uvalue)
|
|
|
|
{
|
|
|
|
struct drgn_type_enumerator *enumerator =
|
|
|
|
drgn_type_enumerator_vector_append_entry(&builder->enumerators);
|
|
|
|
if (!enumerator)
|
|
|
|
return &drgn_enomem;
|
|
|
|
enumerator->name = name;
|
|
|
|
enumerator->uvalue = uvalue;
|
|
|
|
return NULL;
|
2019-11-15 01:12:47 +00:00
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_error *drgn_enum_type_create(struct drgn_enum_type_builder *builder,
|
|
|
|
const char *tag,
|
|
|
|
struct drgn_type *compatible_type,
|
|
|
|
const struct drgn_language *lang,
|
|
|
|
struct drgn_type **ret)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
2021-01-07 01:13:56 +00:00
|
|
|
struct drgn_error *err;
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
if (drgn_type_program(compatible_type) != builder->prog) {
|
|
|
|
return drgn_error_create(DRGN_ERROR_INVALID_ARGUMENT,
|
|
|
|
"type is from different program");
|
|
|
|
}
|
|
|
|
if (drgn_type_kind(compatible_type) != DRGN_TYPE_INT) {
|
|
|
|
return drgn_error_create(DRGN_ERROR_TYPE,
|
|
|
|
"compatible type of enum type must be integer type");
|
|
|
|
}
|
|
|
|
|
2021-01-02 09:41:24 +00:00
|
|
|
if (!builder->enumerators.size) {
|
|
|
|
struct drgn_type key = {
|
|
|
|
{
|
|
|
|
.kind = DRGN_TYPE_ENUM,
|
|
|
|
.is_complete = true,
|
|
|
|
.primitive = DRGN_NOT_PRIMITIVE_TYPE,
|
|
|
|
.tag = tag,
|
|
|
|
.type = compatible_type,
|
|
|
|
.program = builder->prog,
|
|
|
|
.language =
|
|
|
|
lang ? lang : drgn_program_language(builder->prog),
|
|
|
|
}
|
|
|
|
};
|
2021-01-07 01:13:56 +00:00
|
|
|
err = find_or_create_type(&key, ret);
|
|
|
|
if (!err)
|
|
|
|
drgn_type_enumerator_vector_deinit(&builder->enumerators);
|
|
|
|
return err;
|
2021-01-02 09:41:24 +00:00
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_type *type = malloc(sizeof(*type));
|
|
|
|
if (!type)
|
|
|
|
return &drgn_enomem;
|
|
|
|
if (!drgn_typep_vector_append(&builder->prog->created_types, &type)) {
|
|
|
|
free(type);
|
|
|
|
return &drgn_enomem;
|
|
|
|
}
|
|
|
|
|
|
|
|
drgn_type_enumerator_vector_shrink_to_fit(&builder->enumerators);
|
|
|
|
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
type->_private.kind = DRGN_TYPE_ENUM;
|
|
|
|
type->_private.is_complete = true;
|
2019-04-20 00:05:19 +01:00
|
|
|
type->_private.primitive = DRGN_NOT_PRIMITIVE_TYPE;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
type->_private.tag = tag;
|
|
|
|
type->_private.type = compatible_type;
|
|
|
|
type->_private.qualifiers = 0;
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
type->_private.enumerators = builder->enumerators.data;
|
|
|
|
type->_private.num_enumerators = builder->enumerators.size;
|
|
|
|
type->_private.program = builder->prog;
|
|
|
|
type->_private.language =
|
|
|
|
lang ? lang : drgn_program_language(builder->prog);
|
|
|
|
*ret = type;
|
|
|
|
return NULL;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_error *
|
|
|
|
drgn_incomplete_enum_type_create(struct drgn_program *prog, const char *tag,
|
|
|
|
const struct drgn_language *lang,
|
|
|
|
struct drgn_type **ret)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_type key = {
|
|
|
|
{
|
|
|
|
.kind = DRGN_TYPE_ENUM,
|
|
|
|
.is_complete = false,
|
|
|
|
.primitive = DRGN_NOT_PRIMITIVE_TYPE,
|
|
|
|
.tag = tag,
|
|
|
|
.program = prog,
|
|
|
|
.language = lang ? lang : drgn_program_language(prog),
|
|
|
|
}
|
|
|
|
};
|
|
|
|
return find_or_create_type(&key, ret);
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_error *
|
|
|
|
drgn_typedef_type_create(struct drgn_program *prog, const char *name,
|
|
|
|
struct drgn_qualified_type aliased_type,
|
|
|
|
const struct drgn_language *lang,
|
|
|
|
struct drgn_type **ret)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
if (drgn_type_program(aliased_type.type) != prog) {
|
|
|
|
return drgn_error_create(DRGN_ERROR_INVALID_ARGUMENT,
|
|
|
|
"type is from different program");
|
|
|
|
}
|
|
|
|
|
|
|
|
enum drgn_primitive_type primitive;
|
2019-04-20 00:05:19 +01:00
|
|
|
if (strcmp(name, "size_t") == 0)
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
primitive = DRGN_C_TYPE_SIZE_T;
|
2019-04-20 00:05:19 +01:00
|
|
|
else if (strcmp(name, "ptrdiff_t") == 0)
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
primitive = DRGN_C_TYPE_PTRDIFF_T;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
else
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
primitive = DRGN_NOT_PRIMITIVE_TYPE;
|
|
|
|
|
|
|
|
struct drgn_type key = {
|
|
|
|
{
|
|
|
|
.kind = DRGN_TYPE_TYPEDEF,
|
|
|
|
.is_complete = drgn_type_is_complete(aliased_type.type),
|
|
|
|
.primitive = primitive,
|
|
|
|
.name = name,
|
|
|
|
.type = aliased_type.type,
|
|
|
|
.qualifiers = aliased_type.qualifiers,
|
|
|
|
.program = prog,
|
|
|
|
.language = lang ? lang : drgn_program_language(prog),
|
|
|
|
}
|
|
|
|
};
|
|
|
|
return find_or_create_type(&key, ret);
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_error *
|
|
|
|
drgn_pointer_type_create(struct drgn_program *prog,
|
|
|
|
struct drgn_qualified_type referenced_type,
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
uint64_t size, enum drgn_byte_order byte_order,
|
|
|
|
const struct drgn_language *lang,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_type **ret)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
struct drgn_error *err;
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
if (drgn_type_program(referenced_type.type) != prog) {
|
|
|
|
return drgn_error_create(DRGN_ERROR_INVALID_ARGUMENT,
|
|
|
|
"type is from different program");
|
|
|
|
}
|
|
|
|
|
|
|
|
struct drgn_type key = {
|
|
|
|
{
|
|
|
|
.kind = DRGN_TYPE_POINTER,
|
|
|
|
.is_complete = true,
|
|
|
|
.primitive = DRGN_NOT_PRIMITIVE_TYPE,
|
|
|
|
.size = size,
|
|
|
|
.type = referenced_type.type,
|
|
|
|
.qualifiers = referenced_type.qualifiers,
|
|
|
|
.program = prog,
|
|
|
|
.language = lang ? lang : drgn_program_language(prog),
|
|
|
|
}
|
|
|
|
};
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
err = drgn_byte_order_to_little_endian(prog, byte_order,
|
|
|
|
&key._private.little_endian);
|
|
|
|
if (err)
|
|
|
|
return err;
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
return find_or_create_type(&key, ret);
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_error *
|
|
|
|
drgn_array_type_create(struct drgn_program *prog,
|
|
|
|
struct drgn_qualified_type element_type,
|
|
|
|
uint64_t length, const struct drgn_language *lang,
|
|
|
|
struct drgn_type **ret)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
if (drgn_type_program(element_type.type) != prog) {
|
|
|
|
return drgn_error_create(DRGN_ERROR_INVALID_ARGUMENT,
|
|
|
|
"type is from different program");
|
|
|
|
}
|
|
|
|
|
|
|
|
struct drgn_type key = {
|
|
|
|
{
|
|
|
|
.kind = DRGN_TYPE_ARRAY,
|
|
|
|
.is_complete = true,
|
|
|
|
.primitive = DRGN_NOT_PRIMITIVE_TYPE,
|
|
|
|
.length = length,
|
|
|
|
.type = element_type.type,
|
|
|
|
.qualifiers = element_type.qualifiers,
|
|
|
|
.program = prog,
|
|
|
|
.language = lang ? lang : drgn_program_language(prog),
|
|
|
|
}
|
|
|
|
};
|
|
|
|
return find_or_create_type(&key, ret);
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_error *
|
|
|
|
drgn_incomplete_array_type_create(struct drgn_program *prog,
|
|
|
|
struct drgn_qualified_type element_type,
|
|
|
|
const struct drgn_language *lang,
|
|
|
|
struct drgn_type **ret)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
if (drgn_type_program(element_type.type) != prog) {
|
|
|
|
return drgn_error_create(DRGN_ERROR_INVALID_ARGUMENT,
|
|
|
|
"type is from different program");
|
|
|
|
}
|
|
|
|
|
|
|
|
struct drgn_type key = {
|
|
|
|
{
|
|
|
|
.kind = DRGN_TYPE_ARRAY,
|
|
|
|
.is_complete = false,
|
|
|
|
.primitive = DRGN_NOT_PRIMITIVE_TYPE,
|
|
|
|
.type = element_type.type,
|
|
|
|
.qualifiers = element_type.qualifiers,
|
|
|
|
.program = prog,
|
|
|
|
.language = lang ? lang : drgn_program_language(prog),
|
|
|
|
}
|
|
|
|
};
|
|
|
|
return find_or_create_type(&key, ret);
|
|
|
|
}
|
|
|
|
|
|
|
|
DEFINE_VECTOR_FUNCTIONS(drgn_type_parameter_vector)
|
|
|
|
|
|
|
|
void drgn_function_type_builder_init(struct drgn_function_type_builder *builder,
|
|
|
|
struct drgn_program *prog)
|
|
|
|
{
|
2021-01-09 01:28:27 +00:00
|
|
|
drgn_template_parameters_builder_init(&builder->template_builder, prog);
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
drgn_type_parameter_vector_init(&builder->parameters);
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
void
|
|
|
|
drgn_function_type_builder_deinit(struct drgn_function_type_builder *builder)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
for (size_t i = 0; i < builder->parameters.size; i++)
|
2020-12-18 19:01:29 +00:00
|
|
|
drgn_lazy_object_deinit(&builder->parameters.data[i].default_argument);
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
drgn_type_parameter_vector_deinit(&builder->parameters);
|
2021-01-09 01:28:27 +00:00
|
|
|
drgn_template_parameters_builder_deinit(&builder->template_builder);
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
struct drgn_error *
|
|
|
|
drgn_function_type_builder_add_parameter(struct drgn_function_type_builder *builder,
|
2020-12-18 19:01:29 +00:00
|
|
|
const union drgn_lazy_object *default_argument,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
const char *name)
|
|
|
|
{
|
2021-01-09 01:28:27 +00:00
|
|
|
struct drgn_error *err =
|
|
|
|
drgn_lazy_object_check_prog(default_argument,
|
|
|
|
builder->template_builder.prog);
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
struct drgn_type_parameter *parameter =
|
|
|
|
drgn_type_parameter_vector_append_entry(&builder->parameters);
|
|
|
|
if (!parameter)
|
|
|
|
return &drgn_enomem;
|
2020-12-18 19:01:29 +00:00
|
|
|
parameter->default_argument = *default_argument;
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
parameter->name = name;
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
struct drgn_error *
|
|
|
|
drgn_function_type_create(struct drgn_function_type_builder *builder,
|
|
|
|
struct drgn_qualified_type return_type,
|
|
|
|
bool is_variadic, const struct drgn_language *lang,
|
|
|
|
struct drgn_type **ret)
|
|
|
|
{
|
2021-01-07 01:13:56 +00:00
|
|
|
struct drgn_error *err;
|
2021-01-09 01:28:27 +00:00
|
|
|
struct drgn_program *prog = builder->template_builder.prog;
|
2021-01-07 01:13:56 +00:00
|
|
|
|
2021-01-09 01:28:27 +00:00
|
|
|
if (drgn_type_program(return_type.type) != prog) {
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
return drgn_error_create(DRGN_ERROR_INVALID_ARGUMENT,
|
|
|
|
"type is from different program");
|
|
|
|
}
|
|
|
|
|
2021-01-09 01:28:27 +00:00
|
|
|
if (!builder->parameters.size &&
|
|
|
|
!builder->template_builder.parameters.size) {
|
2021-01-02 09:41:24 +00:00
|
|
|
struct drgn_type key = {
|
|
|
|
{
|
|
|
|
.kind = DRGN_TYPE_FUNCTION,
|
|
|
|
.is_complete = true,
|
|
|
|
.primitive = DRGN_NOT_PRIMITIVE_TYPE,
|
|
|
|
.type = return_type.type,
|
|
|
|
.qualifiers = return_type.qualifiers,
|
|
|
|
.is_variadic = is_variadic,
|
2021-01-09 01:28:27 +00:00
|
|
|
.program = prog,
|
2021-01-02 09:41:24 +00:00
|
|
|
.language =
|
2021-01-09 01:28:27 +00:00
|
|
|
lang ? lang : drgn_program_language(prog),
|
2021-01-02 09:41:24 +00:00
|
|
|
}
|
|
|
|
};
|
2021-01-07 01:13:56 +00:00
|
|
|
err = find_or_create_type(&key, ret);
|
|
|
|
if (!err)
|
|
|
|
drgn_type_parameter_vector_deinit(&builder->parameters);
|
|
|
|
return err;
|
2021-01-02 09:41:24 +00:00
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_type *type = malloc(sizeof(*type));
|
|
|
|
if (!type)
|
|
|
|
return &drgn_enomem;
|
2021-01-09 01:28:27 +00:00
|
|
|
if (!drgn_typep_vector_append(&prog->created_types, &type)) {
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
free(type);
|
|
|
|
return &drgn_enomem;
|
|
|
|
}
|
|
|
|
|
|
|
|
drgn_type_parameter_vector_shrink_to_fit(&builder->parameters);
|
2021-01-09 01:28:27 +00:00
|
|
|
drgn_type_template_parameter_vector_shrink_to_fit(&builder->template_builder.parameters);
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
type->_private.kind = DRGN_TYPE_FUNCTION;
|
|
|
|
type->_private.is_complete = true;
|
2019-04-20 00:05:19 +01:00
|
|
|
type->_private.primitive = DRGN_NOT_PRIMITIVE_TYPE;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
type->_private.type = return_type.type;
|
|
|
|
type->_private.qualifiers = return_type.qualifiers;
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
type->_private.parameters = builder->parameters.data;
|
|
|
|
type->_private.num_parameters = builder->parameters.size;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
type->_private.is_variadic = is_variadic;
|
2021-01-09 01:28:27 +00:00
|
|
|
type->_private.template_parameters =
|
|
|
|
builder->template_builder.parameters.data;
|
|
|
|
type->_private.num_template_parameters =
|
|
|
|
builder->template_builder.parameters.size;
|
|
|
|
type->_private.program = prog;
|
|
|
|
type->_private.language = lang ? lang : drgn_program_language(prog);
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
*ret = type;
|
|
|
|
return NULL;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
static struct drgn_error *
|
|
|
|
drgn_type_with_byte_order_impl(struct drgn_type **type,
|
|
|
|
struct drgn_type **underlying_type,
|
|
|
|
enum drgn_byte_order byte_order)
|
|
|
|
{
|
|
|
|
struct drgn_error *err;
|
|
|
|
switch (drgn_type_kind(*type)) {
|
|
|
|
case DRGN_TYPE_INT:
|
|
|
|
err = drgn_int_type_create(drgn_type_program(*type),
|
|
|
|
drgn_type_name(*type),
|
|
|
|
drgn_type_size(*type),
|
|
|
|
drgn_type_is_signed(*type),
|
|
|
|
byte_order,
|
|
|
|
drgn_type_language(*type), type);
|
|
|
|
if (!err)
|
|
|
|
*underlying_type = *type;
|
|
|
|
return err;
|
|
|
|
case DRGN_TYPE_BOOL:
|
|
|
|
err = drgn_bool_type_create(drgn_type_program(*type),
|
|
|
|
drgn_type_name(*type),
|
|
|
|
drgn_type_size(*type), byte_order,
|
|
|
|
drgn_type_language(*type), type);
|
|
|
|
if (!err)
|
|
|
|
*underlying_type = *type;
|
|
|
|
return err;
|
|
|
|
case DRGN_TYPE_FLOAT:
|
|
|
|
err = drgn_float_type_create(drgn_type_program(*type),
|
|
|
|
drgn_type_name(*type),
|
|
|
|
drgn_type_size(*type), byte_order,
|
|
|
|
drgn_type_language(*type), type);
|
|
|
|
if (!err)
|
|
|
|
*underlying_type = *type;
|
|
|
|
return err;
|
|
|
|
case DRGN_TYPE_POINTER:
|
|
|
|
err = drgn_pointer_type_create(drgn_type_program(*type),
|
|
|
|
drgn_type_type(*type),
|
|
|
|
drgn_type_size(*type),
|
|
|
|
byte_order,
|
|
|
|
drgn_type_language(*type), type);
|
|
|
|
if (!err)
|
|
|
|
*underlying_type = *type;
|
|
|
|
return err;
|
|
|
|
case DRGN_TYPE_TYPEDEF: {
|
|
|
|
struct drgn_qualified_type aliased_type = drgn_type_type(*type);
|
|
|
|
err = drgn_type_with_byte_order_impl(&aliased_type.type,
|
|
|
|
underlying_type,
|
|
|
|
byte_order);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
return drgn_typedef_type_create(drgn_type_program(*type),
|
|
|
|
drgn_type_name(*type),
|
|
|
|
aliased_type,
|
|
|
|
drgn_type_language(*type),
|
|
|
|
type);
|
|
|
|
}
|
|
|
|
case DRGN_TYPE_ENUM: {
|
|
|
|
assert(drgn_type_is_complete(*type));
|
|
|
|
struct drgn_type *compatible_type = drgn_type_type(*type).type;
|
|
|
|
struct drgn_type *unused;
|
|
|
|
err = drgn_type_with_byte_order_impl(&compatible_type, &unused,
|
|
|
|
byte_order);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
struct drgn_enum_type_builder builder;
|
|
|
|
drgn_enum_type_builder_init(&builder,
|
|
|
|
drgn_type_program(*type));
|
|
|
|
size_t num_enumerators =
|
|
|
|
drgn_type_num_enumerators(*type);
|
2021-04-02 18:56:37 +01:00
|
|
|
if (num_enumerators) {
|
|
|
|
if (!drgn_type_enumerator_vector_reserve(&builder.enumerators,
|
|
|
|
num_enumerators)) {
|
|
|
|
drgn_enum_type_builder_deinit(&builder);
|
|
|
|
return &drgn_enomem;
|
|
|
|
}
|
|
|
|
memcpy(&builder.enumerators.data,
|
|
|
|
drgn_type_enumerators(*type),
|
|
|
|
num_enumerators * sizeof(struct drgn_type_enumerator));
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
}
|
|
|
|
err = drgn_enum_type_create(&builder, drgn_type_tag(*type),
|
|
|
|
compatible_type,
|
|
|
|
drgn_type_language(*type), type);
|
|
|
|
if (err) {
|
|
|
|
drgn_enum_type_builder_deinit(&builder);
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
*underlying_type = *type;
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
default:
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
struct drgn_error *
|
|
|
|
drgn_type_with_byte_order(struct drgn_type **type,
|
|
|
|
struct drgn_type **underlying_type,
|
|
|
|
enum drgn_byte_order byte_order)
|
|
|
|
{
|
|
|
|
struct drgn_error *err;
|
|
|
|
bool type_little_endian;
|
|
|
|
if (drgn_type_has_little_endian(*underlying_type)) {
|
|
|
|
type_little_endian =
|
|
|
|
drgn_type_little_endian(*underlying_type);
|
|
|
|
} else if (drgn_type_kind(*underlying_type) == DRGN_TYPE_ENUM &&
|
|
|
|
drgn_type_is_complete(*underlying_type)) {
|
|
|
|
type_little_endian =
|
|
|
|
drgn_type_little_endian(drgn_type_type(*underlying_type).type);
|
|
|
|
} else {
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
bool little_endian;
|
|
|
|
err = drgn_byte_order_to_little_endian(drgn_type_program(*underlying_type),
|
|
|
|
byte_order, &little_endian);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
if (type_little_endian == little_endian)
|
|
|
|
return NULL;
|
|
|
|
return drgn_type_with_byte_order_impl(type, underlying_type,
|
|
|
|
drgn_byte_order_from_little_endian(little_endian));
|
|
|
|
}
|
|
|
|
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
LIBDRGN_PUBLIC struct drgn_error *
|
2019-11-28 03:27:14 +00:00
|
|
|
drgn_format_type_name(struct drgn_qualified_type qualified_type, char **ret)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
2020-02-26 21:22:51 +00:00
|
|
|
const struct drgn_language *lang = drgn_type_language(qualified_type.type);
|
|
|
|
return lang->format_type_name(qualified_type, ret);
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
LIBDRGN_PUBLIC struct drgn_error *
|
2019-11-28 03:27:14 +00:00
|
|
|
drgn_format_type(struct drgn_qualified_type qualified_type, char **ret)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
2020-02-26 21:22:51 +00:00
|
|
|
const struct drgn_language *lang = drgn_type_language(qualified_type.type);
|
|
|
|
return lang->format_type(qualified_type, ret);
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
bool drgn_type_is_integer(struct drgn_type *type)
|
|
|
|
{
|
|
|
|
switch (drgn_type_kind(type)) {
|
|
|
|
case DRGN_TYPE_INT:
|
|
|
|
case DRGN_TYPE_BOOL:
|
|
|
|
case DRGN_TYPE_ENUM:
|
|
|
|
return true;
|
|
|
|
case DRGN_TYPE_TYPEDEF:
|
|
|
|
return drgn_type_is_integer(drgn_type_type(type).type);
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
bool drgn_type_is_arithmetic(struct drgn_type *type)
|
|
|
|
{
|
|
|
|
switch (drgn_type_kind(type)) {
|
|
|
|
case DRGN_TYPE_INT:
|
|
|
|
case DRGN_TYPE_BOOL:
|
|
|
|
case DRGN_TYPE_FLOAT:
|
|
|
|
case DRGN_TYPE_ENUM:
|
|
|
|
return true;
|
|
|
|
case DRGN_TYPE_TYPEDEF:
|
|
|
|
return drgn_type_is_arithmetic(drgn_type_type(type).type);
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
bool drgn_type_is_scalar(struct drgn_type *type)
|
|
|
|
{
|
|
|
|
switch (drgn_type_kind(type)) {
|
|
|
|
case DRGN_TYPE_INT:
|
|
|
|
case DRGN_TYPE_BOOL:
|
|
|
|
case DRGN_TYPE_FLOAT:
|
|
|
|
case DRGN_TYPE_ENUM:
|
|
|
|
case DRGN_TYPE_POINTER:
|
|
|
|
return true;
|
|
|
|
case DRGN_TYPE_TYPEDEF:
|
|
|
|
return drgn_type_is_scalar(drgn_type_type(type).type);
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
LIBDRGN_PUBLIC struct drgn_error *drgn_type_sizeof(struct drgn_type *type,
|
|
|
|
uint64_t *ret)
|
|
|
|
{
|
|
|
|
struct drgn_error *err;
|
2019-11-15 01:12:47 +00:00
|
|
|
enum drgn_type_kind kind = drgn_type_kind(type);
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
|
2019-11-15 01:12:47 +00:00
|
|
|
if (!drgn_type_is_complete(type)) {
|
|
|
|
return drgn_error_format(DRGN_ERROR_TYPE,
|
|
|
|
"cannot get size of incomplete %s type",
|
|
|
|
drgn_type_kind_spelling[kind]);
|
|
|
|
}
|
2020-12-04 19:20:00 +00:00
|
|
|
SWITCH_ENUM(kind,
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
case DRGN_TYPE_INT:
|
|
|
|
case DRGN_TYPE_BOOL:
|
|
|
|
case DRGN_TYPE_FLOAT:
|
|
|
|
case DRGN_TYPE_POINTER:
|
|
|
|
*ret = drgn_type_size(type);
|
|
|
|
return NULL;
|
|
|
|
case DRGN_TYPE_STRUCT:
|
|
|
|
case DRGN_TYPE_UNION:
|
2019-11-15 01:12:47 +00:00
|
|
|
case DRGN_TYPE_CLASS:
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
*ret = drgn_type_size(type);
|
|
|
|
return NULL;
|
|
|
|
case DRGN_TYPE_ENUM:
|
|
|
|
case DRGN_TYPE_TYPEDEF:
|
|
|
|
return drgn_type_sizeof(drgn_type_type(type).type, ret);
|
|
|
|
case DRGN_TYPE_ARRAY:
|
|
|
|
err = drgn_type_sizeof(drgn_type_type(type).type, ret);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
if (__builtin_mul_overflow(*ret, drgn_type_length(type), ret)) {
|
|
|
|
return drgn_error_create(DRGN_ERROR_OVERFLOW,
|
|
|
|
"type size is too large");
|
|
|
|
}
|
|
|
|
return NULL;
|
|
|
|
case DRGN_TYPE_VOID:
|
|
|
|
return drgn_error_create(DRGN_ERROR_TYPE,
|
|
|
|
"cannot get size of void type");
|
|
|
|
case DRGN_TYPE_FUNCTION:
|
|
|
|
return drgn_error_create(DRGN_ERROR_TYPE,
|
|
|
|
"cannot get size of function type");
|
2020-12-04 19:20:00 +00:00
|
|
|
)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
struct drgn_error *drgn_type_bit_size(struct drgn_type *type, uint64_t *ret)
|
|
|
|
{
|
|
|
|
struct drgn_error *err;
|
|
|
|
|
|
|
|
err = drgn_type_sizeof(type, ret);
|
|
|
|
if (err)
|
|
|
|
return err;
|
2019-03-30 06:43:21 +00:00
|
|
|
if (__builtin_mul_overflow(*ret, 8U, ret)) {
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
return drgn_error_create(DRGN_ERROR_OVERFLOW,
|
|
|
|
"type bit size is too large");
|
|
|
|
}
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2020-12-04 20:01:29 +00:00
|
|
|
enum drgn_object_encoding drgn_type_object_encoding(struct drgn_type *type)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
{
|
2020-12-04 19:20:00 +00:00
|
|
|
SWITCH_ENUM(drgn_type_kind(type),
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
case DRGN_TYPE_INT:
|
2020-12-04 20:01:29 +00:00
|
|
|
return (drgn_type_is_signed(type) ?
|
|
|
|
DRGN_OBJECT_ENCODING_SIGNED :
|
|
|
|
DRGN_OBJECT_ENCODING_UNSIGNED);
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
case DRGN_TYPE_BOOL:
|
|
|
|
case DRGN_TYPE_POINTER:
|
2020-12-04 20:01:29 +00:00
|
|
|
return DRGN_OBJECT_ENCODING_UNSIGNED;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
case DRGN_TYPE_FLOAT:
|
2020-12-04 20:01:29 +00:00
|
|
|
return DRGN_OBJECT_ENCODING_FLOAT;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
case DRGN_TYPE_STRUCT:
|
|
|
|
case DRGN_TYPE_UNION:
|
2019-11-15 01:12:47 +00:00
|
|
|
case DRGN_TYPE_CLASS:
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
case DRGN_TYPE_ARRAY:
|
2020-12-04 20:01:29 +00:00
|
|
|
return (drgn_type_is_complete(type) ?
|
|
|
|
DRGN_OBJECT_ENCODING_BUFFER :
|
|
|
|
DRGN_OBJECT_ENCODING_INCOMPLETE_BUFFER);
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
case DRGN_TYPE_ENUM:
|
|
|
|
if (!drgn_type_is_complete(type))
|
2020-12-04 20:01:29 +00:00
|
|
|
return DRGN_OBJECT_ENCODING_INCOMPLETE_INTEGER;
|
2022-10-05 07:36:01 +01:00
|
|
|
fallthrough;
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
case DRGN_TYPE_TYPEDEF:
|
2020-12-04 20:01:29 +00:00
|
|
|
return drgn_type_object_encoding(drgn_type_type(type).type);
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
case DRGN_TYPE_VOID:
|
|
|
|
case DRGN_TYPE_FUNCTION:
|
2020-12-04 20:01:29 +00:00
|
|
|
return DRGN_OBJECT_ENCODING_NONE;
|
2020-12-04 19:20:00 +00:00
|
|
|
)
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
struct drgn_error *drgn_type_error(const char *format, struct drgn_type *type)
|
|
|
|
{
|
|
|
|
struct drgn_qualified_type qualified_type = { type };
|
|
|
|
|
|
|
|
return drgn_qualified_type_error(format, qualified_type);
|
|
|
|
}
|
|
|
|
|
|
|
|
struct drgn_error *
|
|
|
|
drgn_qualified_type_error(const char *format,
|
|
|
|
struct drgn_qualified_type qualified_type)
|
|
|
|
{
|
|
|
|
struct drgn_error *err;
|
|
|
|
char *name;
|
|
|
|
|
2019-11-28 03:27:14 +00:00
|
|
|
err = drgn_format_type_name(qualified_type, &name);
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
err = drgn_error_format(DRGN_ERROR_TYPE, format, name);
|
|
|
|
free(name);
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
|
|
|
struct drgn_error *drgn_error_incomplete_type(const char *format,
|
|
|
|
struct drgn_type *type)
|
|
|
|
{
|
|
|
|
switch (drgn_type_kind(drgn_underlying_type(type))) {
|
|
|
|
case DRGN_TYPE_STRUCT:
|
|
|
|
return drgn_error_format(DRGN_ERROR_TYPE, format,
|
|
|
|
"incomplete structure");
|
|
|
|
case DRGN_TYPE_UNION:
|
|
|
|
return drgn_error_format(DRGN_ERROR_TYPE, format,
|
|
|
|
"incomplete union");
|
2019-11-15 01:12:47 +00:00
|
|
|
case DRGN_TYPE_CLASS:
|
|
|
|
return drgn_error_format(DRGN_ERROR_TYPE, format,
|
|
|
|
"incomplete class");
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
case DRGN_TYPE_ENUM:
|
|
|
|
return drgn_error_format(DRGN_ERROR_TYPE, format,
|
|
|
|
"incomplete enumerated");
|
|
|
|
case DRGN_TYPE_ARRAY:
|
|
|
|
return drgn_error_format(DRGN_ERROR_TYPE, format,
|
|
|
|
"incomplete array");
|
|
|
|
case DRGN_TYPE_FUNCTION:
|
|
|
|
return drgn_error_format(DRGN_ERROR_TYPE, format, "function");
|
|
|
|
case DRGN_TYPE_VOID:
|
|
|
|
return drgn_error_format(DRGN_ERROR_TYPE, format, "void");
|
|
|
|
default:
|
2020-05-07 23:02:33 +01:00
|
|
|
UNREACHABLE();
|
Rewrite drgn core in C
The current mixed Python/C implementation works well, but it has a
couple of important limitations:
- It's too slow for some common use cases, like iterating over large
data structures.
- It can't be reused in utilities written in other languages.
This replaces the internals with a new library written in C, libdrgn. It
includes Python bindings with mostly the same public interface as
before, with some important improvements:
- Types are now represented by a single Type class rather than the messy
polymorphism in the Python implementation.
- Qualifiers are a bitmask instead of a set of strings.
- Bit fields are not considered a separate type.
- The lvalue/rvalue terminology is replaced with reference/value.
- Structure, union, and array values are better supported.
- Function objects are supported.
- Program distinguishes between lookups of variables, constants, and
functions.
The C rewrite is about 6x as fast as the original Python when using the
Python bindings, and about 8x when using the C API directly.
Currently, the exposed API in C is fairly conservative. In the future,
the memory reader, type index, and object index APIs will probably be
exposed for more flexibility.
2019-03-22 23:27:46 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2020-04-23 00:23:26 +01:00
|
|
|
void drgn_program_init_types(struct drgn_program *prog)
|
|
|
|
{
|
2021-08-24 01:07:26 +01:00
|
|
|
for (size_t i = 0; i < array_size(prog->void_types); i++) {
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_type *type = &prog->void_types[i];
|
|
|
|
type->_private.kind = DRGN_TYPE_VOID;
|
|
|
|
type->_private.is_complete = false;
|
|
|
|
type->_private.primitive = DRGN_C_TYPE_VOID;
|
|
|
|
type->_private.program = prog;
|
2022-02-16 20:43:36 +00:00
|
|
|
type->_private.language = drgn_languages[i];
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
}
|
|
|
|
drgn_dedupe_type_set_init(&prog->dedupe_types);
|
|
|
|
drgn_typep_vector_init(&prog->created_types);
|
2020-04-23 00:23:26 +01:00
|
|
|
drgn_member_map_init(&prog->members);
|
|
|
|
drgn_type_set_init(&prog->members_cached);
|
|
|
|
}
|
|
|
|
|
|
|
|
void drgn_program_deinit_types(struct drgn_program *prog)
|
|
|
|
{
|
|
|
|
drgn_member_map_deinit(&prog->members);
|
|
|
|
drgn_type_set_deinit(&prog->members_cached);
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
|
|
|
|
for (size_t i = 0; i < prog->created_types.size; i++) {
|
|
|
|
struct drgn_type *type = prog->created_types.data[i];
|
|
|
|
if (drgn_type_has_members(type)) {
|
|
|
|
struct drgn_type_member *members =
|
|
|
|
drgn_type_members(type);
|
|
|
|
size_t num_members = drgn_type_num_members(type);
|
|
|
|
for (size_t j = 0; j < num_members; j++)
|
2020-12-18 19:01:29 +00:00
|
|
|
drgn_lazy_object_deinit(&members[j].object);
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
free(members);
|
|
|
|
}
|
|
|
|
if (drgn_type_has_enumerators(type))
|
|
|
|
free(drgn_type_enumerators(type));
|
|
|
|
if (drgn_type_has_parameters(type)) {
|
|
|
|
struct drgn_type_parameter *parameters =
|
|
|
|
drgn_type_parameters(type);
|
|
|
|
size_t num_parameters = drgn_type_num_parameters(type);
|
|
|
|
for (size_t j = 0; j < num_parameters; j++)
|
2020-12-18 19:01:29 +00:00
|
|
|
drgn_lazy_object_deinit(¶meters[j].default_argument);
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
free(parameters);
|
|
|
|
}
|
|
|
|
free(type);
|
|
|
|
}
|
|
|
|
drgn_typep_vector_deinit(&prog->created_types);
|
|
|
|
|
|
|
|
for (struct drgn_dedupe_type_set_iterator it =
|
|
|
|
drgn_dedupe_type_set_first(&prog->dedupe_types);
|
|
|
|
it.entry; it = drgn_dedupe_type_set_next(it))
|
|
|
|
free(*it.entry);
|
|
|
|
drgn_dedupe_type_set_deinit(&prog->dedupe_types);
|
2020-04-23 00:23:26 +01:00
|
|
|
|
|
|
|
struct drgn_type_finder *finder = prog->type_finders;
|
|
|
|
while (finder) {
|
|
|
|
struct drgn_type_finder *next = finder->next;
|
|
|
|
free(finder);
|
|
|
|
finder = next;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
LIBDRGN_PUBLIC struct drgn_error *
|
|
|
|
drgn_program_add_type_finder(struct drgn_program *prog, drgn_type_find_fn fn,
|
|
|
|
void *arg)
|
|
|
|
{
|
|
|
|
struct drgn_type_finder *finder = malloc(sizeof(*finder));
|
|
|
|
if (!finder)
|
|
|
|
return &drgn_enomem;
|
|
|
|
finder->fn = fn;
|
|
|
|
finder->arg = arg;
|
|
|
|
finder->next = prog->type_finders;
|
|
|
|
prog->type_finders = finder;
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
struct drgn_error *
|
|
|
|
drgn_program_find_type_impl(struct drgn_program *prog,
|
|
|
|
enum drgn_type_kind kind, const char *name,
|
|
|
|
size_t name_len, const char *filename,
|
|
|
|
struct drgn_qualified_type *ret)
|
|
|
|
{
|
|
|
|
struct drgn_type_finder *finder = prog->type_finders;
|
|
|
|
while (finder) {
|
|
|
|
struct drgn_error *err =
|
|
|
|
finder->fn(kind, name, name_len, filename, finder->arg,
|
|
|
|
ret);
|
|
|
|
if (!err) {
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
if (drgn_type_program(ret->type) != prog) {
|
|
|
|
return drgn_error_create(DRGN_ERROR_INVALID_ARGUMENT,
|
|
|
|
"type find callback returned type from wrong program");
|
|
|
|
}
|
2020-04-23 00:23:26 +01:00
|
|
|
if (drgn_type_kind(ret->type) != kind) {
|
|
|
|
return drgn_error_create(DRGN_ERROR_TYPE,
|
|
|
|
"type find callback returned wrong kind of type");
|
|
|
|
}
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
if (err != &drgn_not_found)
|
|
|
|
return err;
|
|
|
|
finder = finder->next;
|
|
|
|
}
|
|
|
|
return &drgn_not_found;
|
|
|
|
}
|
|
|
|
|
|
|
|
LIBDRGN_PUBLIC struct drgn_error *
|
|
|
|
drgn_program_find_type(struct drgn_program *prog, const char *name,
|
|
|
|
const char *filename, struct drgn_qualified_type *ret)
|
|
|
|
{
|
|
|
|
struct drgn_error *err;
|
2022-02-18 06:18:53 +00:00
|
|
|
const struct drgn_language *lang = drgn_program_language(prog);
|
|
|
|
err = lang->find_type(lang, prog, name, filename, ret);
|
2020-04-23 00:23:26 +01:00
|
|
|
if (err != &drgn_not_found)
|
|
|
|
return err;
|
|
|
|
|
|
|
|
if (filename) {
|
|
|
|
return drgn_error_format(DRGN_ERROR_LOOKUP,
|
|
|
|
"could not find '%s' in '%s'", name,
|
|
|
|
filename);
|
|
|
|
} else {
|
|
|
|
return drgn_error_format(DRGN_ERROR_LOOKUP,
|
|
|
|
"could not find '%s'", name);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
/*
|
|
|
|
* size_t and ptrdiff_t default to typedefs of whatever integer type matches the
|
|
|
|
* word size.
|
|
|
|
*/
|
|
|
|
static struct drgn_error *
|
|
|
|
default_size_t_or_ptrdiff_t(struct drgn_program *prog,
|
|
|
|
enum drgn_primitive_type type,
|
|
|
|
struct drgn_type **ret)
|
2020-04-23 00:23:26 +01:00
|
|
|
{
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
static const enum drgn_primitive_type integer_types[2][3] = {
|
|
|
|
{
|
|
|
|
DRGN_C_TYPE_UNSIGNED_LONG,
|
|
|
|
DRGN_C_TYPE_UNSIGNED_LONG_LONG,
|
|
|
|
DRGN_C_TYPE_UNSIGNED_INT,
|
|
|
|
},
|
|
|
|
{
|
|
|
|
DRGN_C_TYPE_LONG,
|
|
|
|
DRGN_C_TYPE_LONG_LONG,
|
|
|
|
DRGN_C_TYPE_INT,
|
|
|
|
},
|
|
|
|
};
|
|
|
|
struct drgn_error *err;
|
2021-02-23 22:06:41 +00:00
|
|
|
uint8_t address_size;
|
|
|
|
err = drgn_program_address_size(prog, &address_size);
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
if (err)
|
|
|
|
return err;
|
2021-08-24 01:07:26 +01:00
|
|
|
array_for_each(integer_type,
|
|
|
|
integer_types[type == DRGN_C_TYPE_PTRDIFF_T]) {
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
struct drgn_qualified_type qualified_type;
|
2021-08-24 01:07:26 +01:00
|
|
|
err = drgn_program_find_primitive_type(prog, *integer_type,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
&qualified_type.type);
|
|
|
|
if (err)
|
|
|
|
return err;
|
2021-02-23 22:06:41 +00:00
|
|
|
if (drgn_type_size(qualified_type.type) == address_size) {
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
qualified_type.qualifiers = 0;
|
|
|
|
return drgn_typedef_type_create(prog,
|
|
|
|
drgn_primitive_type_spellings[type][0],
|
|
|
|
qualified_type,
|
|
|
|
&drgn_language_c, ret);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return drgn_error_format(DRGN_ERROR_INVALID_ARGUMENT,
|
|
|
|
"no suitable integer type for %s",
|
|
|
|
drgn_primitive_type_spellings[type][0]);
|
2020-04-23 00:23:26 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
struct drgn_error *
|
|
|
|
drgn_program_find_primitive_type(struct drgn_program *prog,
|
|
|
|
enum drgn_primitive_type type,
|
|
|
|
struct drgn_type **ret)
|
|
|
|
{
|
|
|
|
struct drgn_error *err;
|
|
|
|
struct drgn_qualified_type qualified_type;
|
|
|
|
enum drgn_type_kind kind;
|
|
|
|
const char * const *spellings;
|
|
|
|
size_t i;
|
|
|
|
|
|
|
|
if (prog->primitive_types[type]) {
|
|
|
|
*ret = prog->primitive_types[type];
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
kind = drgn_primitive_type_kind[type];
|
|
|
|
if (kind == DRGN_TYPE_VOID) {
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
*ret = drgn_void_type(prog, &drgn_language_c);
|
2020-04-23 00:23:26 +01:00
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
|
|
|
|
spellings = drgn_primitive_type_spellings[type];
|
|
|
|
for (i = 0; spellings[i]; i++) {
|
|
|
|
err = drgn_program_find_type_impl(prog, kind, spellings[i],
|
|
|
|
strlen(spellings[i]), NULL,
|
|
|
|
&qualified_type);
|
|
|
|
if (!err && drgn_type_primitive(qualified_type.type) == type) {
|
|
|
|
*ret = qualified_type.type;
|
|
|
|
goto out;
|
|
|
|
} else if (err && err != &drgn_not_found) {
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
switch (type) {
|
|
|
|
case DRGN_C_TYPE_CHAR:
|
|
|
|
case DRGN_C_TYPE_SIGNED_CHAR:
|
|
|
|
err = drgn_int_type_create(prog, spellings[0], 1, true,
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
DRGN_PROGRAM_ENDIAN,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
&drgn_language_c, ret);
|
|
|
|
break;
|
|
|
|
case DRGN_C_TYPE_UNSIGNED_CHAR:
|
|
|
|
err = drgn_int_type_create(prog, spellings[0], 1, false,
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
DRGN_PROGRAM_ENDIAN,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
&drgn_language_c, ret);
|
|
|
|
break;
|
|
|
|
case DRGN_C_TYPE_SHORT:
|
|
|
|
err = drgn_int_type_create(prog, spellings[0], 2, true,
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
DRGN_PROGRAM_ENDIAN,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
&drgn_language_c, ret);
|
|
|
|
break;
|
|
|
|
case DRGN_C_TYPE_UNSIGNED_SHORT:
|
|
|
|
err = drgn_int_type_create(prog, spellings[0], 2, false,
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
DRGN_PROGRAM_ENDIAN,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
&drgn_language_c, ret);
|
|
|
|
break;
|
|
|
|
case DRGN_C_TYPE_INT:
|
|
|
|
err = drgn_int_type_create(prog, spellings[0], 4, true,
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
DRGN_PROGRAM_ENDIAN,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
&drgn_language_c, ret);
|
|
|
|
break;
|
|
|
|
case DRGN_C_TYPE_UNSIGNED_INT:
|
|
|
|
err = drgn_int_type_create(prog, spellings[0], 4, false,
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
DRGN_PROGRAM_ENDIAN,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
&drgn_language_c, ret);
|
|
|
|
break;
|
2020-04-23 00:23:26 +01:00
|
|
|
/* long and unsigned long default to the word size. */
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
case DRGN_C_TYPE_LONG:
|
|
|
|
case DRGN_C_TYPE_UNSIGNED_LONG: {
|
2021-02-23 22:06:41 +00:00
|
|
|
uint8_t address_size;
|
|
|
|
err = drgn_program_address_size(prog, &address_size);
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
if (err)
|
|
|
|
break;
|
2021-02-23 22:06:41 +00:00
|
|
|
err = drgn_int_type_create(prog, spellings[0], address_size,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
type == DRGN_C_TYPE_LONG,
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
DRGN_PROGRAM_ENDIAN,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
&drgn_language_c, ret);
|
|
|
|
break;
|
2020-04-23 00:23:26 +01:00
|
|
|
}
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
case DRGN_C_TYPE_LONG_LONG:
|
|
|
|
err = drgn_int_type_create(prog, spellings[0], 8, true,
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
DRGN_PROGRAM_ENDIAN,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
&drgn_language_c, ret);
|
|
|
|
break;
|
|
|
|
case DRGN_C_TYPE_UNSIGNED_LONG_LONG:
|
|
|
|
err = drgn_int_type_create(prog, spellings[0], 8, false,
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
DRGN_PROGRAM_ENDIAN,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
&drgn_language_c, ret);
|
|
|
|
break;
|
|
|
|
case DRGN_C_TYPE_BOOL:
|
|
|
|
err = drgn_bool_type_create(prog, spellings[0], 1,
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
DRGN_PROGRAM_ENDIAN,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
&drgn_language_c, ret);
|
|
|
|
break;
|
|
|
|
case DRGN_C_TYPE_FLOAT:
|
|
|
|
err = drgn_float_type_create(prog, spellings[0], 4,
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
DRGN_PROGRAM_ENDIAN,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
&drgn_language_c, ret);
|
|
|
|
break;
|
|
|
|
case DRGN_C_TYPE_DOUBLE:
|
|
|
|
err = drgn_float_type_create(prog, spellings[0], 8,
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
DRGN_PROGRAM_ENDIAN,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
&drgn_language_c, ret);
|
|
|
|
break;
|
|
|
|
case DRGN_C_TYPE_LONG_DOUBLE:
|
|
|
|
err = drgn_float_type_create(prog, spellings[0], 16,
|
Track byte order in scalar types instead of objects
Currently, reference objects and buffer value objects have a byte order.
However, this doesn't always make sense for a couple of reasons:
- Byte order is only meaningful for scalars. What does it mean for a
struct to be big endian? A struct doesn't have a most or least
significant byte; its scalar members do.
- The DWARF specification allows either types or variables to have a
byte order (DW_AT_endianity). The only producer I could find that uses
this is GCC for the scalar_storage_order type attribute, and it only
uses it for base types, not variables. GDB only seems to use to check
it for base types, as well.
So, remove the byte order from objects, and move it to integer, boolean,
floating-point, and pointer types. This model makes more sense, and it
means that we can get the binary representation of any object now.
The only downside is that we can no longer support a bit offset for
non-scalars, but as far as I can tell, nothing needs that.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2021-02-18 00:13:23 +00:00
|
|
|
DRGN_PROGRAM_ENDIAN,
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
&drgn_language_c, ret);
|
|
|
|
break;
|
|
|
|
case DRGN_C_TYPE_SIZE_T:
|
|
|
|
case DRGN_C_TYPE_PTRDIFF_T:
|
|
|
|
err = default_size_t_or_ptrdiff_t(prog, type, ret);
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
UNREACHABLE();
|
2020-04-23 00:23:26 +01:00
|
|
|
}
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
assert(drgn_type_primitive(*ret) == type);
|
2020-04-23 00:23:26 +01:00
|
|
|
|
|
|
|
out:
|
Associate types with program
I originally envisioned types as dumb descriptors. This mostly works for
C because in C, types are fairly simple. However, even then the
drgn_program_member_info() API is awkward. You should be able to look up
a member directly from a type, but we need the program for caching
purposes. This has also held me back from adding offsetof() or
has_member() APIs.
Things get even messier with C++. C++ template parameters can be objects
(e.g., template <int N>). Such parameters would best be represented by a
drgn object, which we need a drgn program for. Static members are a
similar case.
So, let's reimagine types as being owned by a program. This has a few
parts:
1. In libdrgn, simple types are now created by factory functions,
drgn_foo_type_create().
2. To handle their variable length fields, compound types, enum types,
and function types are constructed with a "builder" API.
3. Simple types are deduplicated.
4. The Python type factory functions are replaced by methods of the
Program class.
5. While we're changing the API, the parameters to pointer_type() and
array_type() are reordered to be more logical (and to allow
pointer_type() to take a default size of None for the program's
default pointer size).
6. Likewise, the type factory methods take qualifiers as a keyword
argument only.
A big part of this change is updating the tests and splitting up large
test cases into smaller ones in a few places.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
2020-07-16 00:34:56 +01:00
|
|
|
prog->primitive_types[type] = *ret;
|
2020-04-23 00:23:26 +01:00
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
static struct drgn_error *
|
2020-12-15 01:58:44 +00:00
|
|
|
drgn_type_cache_members(struct drgn_type *outer_type,
|
|
|
|
struct drgn_type *type, uint64_t bit_offset)
|
2020-04-23 00:23:26 +01:00
|
|
|
{
|
2020-12-15 01:58:44 +00:00
|
|
|
struct drgn_program *prog = drgn_type_program(outer_type);
|
|
|
|
|
2020-04-23 00:23:26 +01:00
|
|
|
if (!drgn_type_has_members(type))
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
struct drgn_type_member *members = drgn_type_members(type);
|
|
|
|
size_t num_members = drgn_type_num_members(type);
|
|
|
|
for (size_t i = 0; i < num_members; i++) {
|
|
|
|
struct drgn_type_member *member = &members[i];
|
|
|
|
if (member->name) {
|
|
|
|
struct drgn_member_map_entry entry = {
|
|
|
|
.key = {
|
|
|
|
.type = outer_type,
|
|
|
|
.name = member->name,
|
|
|
|
.name_len = strlen(member->name),
|
|
|
|
},
|
|
|
|
.value = {
|
2020-12-15 01:58:44 +00:00
|
|
|
.member = member,
|
2020-04-23 00:23:26 +01:00
|
|
|
.bit_offset =
|
|
|
|
bit_offset + member->bit_offset,
|
|
|
|
},
|
|
|
|
};
|
|
|
|
if (drgn_member_map_insert(&prog->members, &entry,
|
|
|
|
NULL) == -1)
|
|
|
|
return &drgn_enomem;
|
|
|
|
} else {
|
|
|
|
struct drgn_qualified_type member_type;
|
|
|
|
struct drgn_error *err = drgn_member_type(member,
|
2020-12-18 22:11:29 +00:00
|
|
|
&member_type,
|
|
|
|
NULL);
|
2020-04-23 00:23:26 +01:00
|
|
|
if (err)
|
|
|
|
return err;
|
2020-12-15 01:58:44 +00:00
|
|
|
err = drgn_type_cache_members(outer_type,
|
|
|
|
member_type.type,
|
|
|
|
bit_offset +
|
|
|
|
member->bit_offset);
|
2020-04-23 00:23:26 +01:00
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2020-12-16 00:35:18 +00:00
|
|
|
LIBDRGN_PUBLIC struct drgn_error *
|
|
|
|
drgn_type_offsetof(struct drgn_type *type, const char *member_designator,
|
|
|
|
uint64_t *ret)
|
|
|
|
{
|
|
|
|
|
|
|
|
struct drgn_error *err;
|
|
|
|
const struct drgn_language *lang = drgn_type_language(type);
|
|
|
|
uint64_t bit_offset;
|
|
|
|
err = lang->bit_offset(drgn_type_program(type), type, member_designator,
|
|
|
|
&bit_offset);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
if (bit_offset % 8) {
|
|
|
|
return drgn_error_format(DRGN_ERROR_INVALID_ARGUMENT,
|
|
|
|
"member is not byte-aligned");
|
|
|
|
}
|
|
|
|
*ret = bit_offset / 8;
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2020-12-15 01:58:44 +00:00
|
|
|
static struct drgn_error *
|
|
|
|
drgn_type_find_member_impl(struct drgn_type *type, const char *member_name,
|
|
|
|
size_t member_name_len,
|
|
|
|
struct drgn_member_value **ret)
|
2020-04-23 00:23:26 +01:00
|
|
|
{
|
2020-12-15 01:58:44 +00:00
|
|
|
struct drgn_program *prog = drgn_type_program(type);
|
2020-04-23 00:23:26 +01:00
|
|
|
const struct drgn_member_key key = {
|
|
|
|
.type = drgn_underlying_type(type),
|
|
|
|
.name = member_name,
|
|
|
|
.name_len = member_name_len,
|
|
|
|
};
|
|
|
|
struct hash_pair hp = drgn_member_map_hash(&key);
|
|
|
|
struct drgn_member_map_iterator it =
|
|
|
|
drgn_member_map_search_hashed(&prog->members, &key, hp);
|
|
|
|
if (it.entry) {
|
|
|
|
*ret = &it.entry->value;
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Cache miss. One of the following is true:
|
|
|
|
*
|
|
|
|
* 1. The type isn't a structure, union, or class, which is a type
|
|
|
|
* error.
|
|
|
|
* 2. The type hasn't been cached, which means we need to cache it and
|
|
|
|
* check again.
|
|
|
|
* 3. The type has already been cached, which means the member doesn't
|
|
|
|
* exist.
|
|
|
|
*/
|
|
|
|
if (!drgn_type_has_members(key.type)) {
|
|
|
|
return drgn_type_error("'%s' is not a structure, union, or class",
|
|
|
|
type);
|
|
|
|
}
|
|
|
|
struct hash_pair cached_hp = drgn_type_set_hash(&key.type);
|
|
|
|
if (drgn_type_set_search_hashed(&prog->members_cached, &key.type,
|
2020-12-15 01:58:44 +00:00
|
|
|
cached_hp).entry) {
|
|
|
|
*ret = NULL;
|
|
|
|
return NULL;
|
|
|
|
}
|
2020-04-23 00:23:26 +01:00
|
|
|
|
2020-12-15 01:58:44 +00:00
|
|
|
struct drgn_error *err = drgn_type_cache_members(key.type, key.type, 0);
|
2020-04-23 00:23:26 +01:00
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
|
|
|
|
if (drgn_type_set_insert_searched(&prog->members_cached, &key.type,
|
|
|
|
cached_hp, NULL) == -1)
|
|
|
|
return &drgn_enomem;
|
|
|
|
|
|
|
|
it = drgn_member_map_search_hashed(&prog->members, &key, hp);
|
|
|
|
if (it.entry) {
|
|
|
|
*ret = &it.entry->value;
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2020-12-15 01:58:44 +00:00
|
|
|
*ret = NULL;
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
LIBDRGN_PUBLIC struct drgn_error *
|
|
|
|
drgn_type_find_member_len(struct drgn_type *type, const char *member_name,
|
|
|
|
size_t member_name_len,
|
|
|
|
struct drgn_type_member **member_ret,
|
|
|
|
uint64_t *bit_offset_ret)
|
|
|
|
{
|
|
|
|
struct drgn_error *err;
|
|
|
|
struct drgn_member_value *member;
|
|
|
|
err = drgn_type_find_member_impl(type, member_name, member_name_len,
|
|
|
|
&member);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
if (!member) {
|
|
|
|
struct drgn_qualified_type qualified_type = { type };
|
|
|
|
char *type_name;
|
|
|
|
err = drgn_format_type_name(qualified_type, &type_name);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
err = drgn_error_format(DRGN_ERROR_LOOKUP,
|
|
|
|
"'%s' has no member '%.*s'",
|
|
|
|
type_name,
|
|
|
|
member_name_len > INT_MAX ?
|
|
|
|
INT_MAX : (int)member_name_len,
|
|
|
|
member_name);
|
|
|
|
free(type_name);
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
*member_ret = member->member;
|
|
|
|
*bit_offset_ret = member->bit_offset;
|
|
|
|
return NULL;
|
2020-04-23 00:23:26 +01:00
|
|
|
}
|
2020-12-15 23:51:11 +00:00
|
|
|
|
|
|
|
LIBDRGN_PUBLIC struct drgn_error *
|
|
|
|
drgn_type_has_member_len(struct drgn_type *type, const char *member_name,
|
|
|
|
size_t member_name_len, bool *ret)
|
|
|
|
{
|
|
|
|
struct drgn_error *err;
|
|
|
|
struct drgn_member_value *member;
|
|
|
|
err = drgn_type_find_member_impl(type, member_name, member_name_len,
|
|
|
|
&member);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
*ret = member != NULL;
|
|
|
|
return NULL;
|
|
|
|
}
|