There is still the problem of handing the strings table to the CTF encoder, but
that will be fixed another day.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
So that we can use the strings in ".strtab" directly, without duplicating them
on the global strings table.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Instead pass thru cu__strings(cu, i) so that we can figure out if the
underlying debugging format handler can do that more efficiently, such as by
looking up directly the ELF section ".strtab".
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Such as signed, etc. This is in preparation for using directly ctf_strings.
Instead of duplicating it in the global strings table.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
In the paste we ass-umed that if cu->dfops != NULL, all the methods would be
there, this ain't so anymore, so check it.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
That is done by adding some new struct debug_fmt_ops methods:
->function__name()
This one, if specified, will be called by function__name(), giving a chance to
formats such as CTF to get this from some other place than the global strings
table. CTF does this by storing GElf_Sym->st_name in function->name, and by
providing a dfops->function__name() that uses function->name as an index into
the .strtab ELF section.
->cu__delete()
This is needed because we can't anymore call ctf__delete at the end of
ctf__load_file, as we will need at least the .strstab ELF section to be
available till we're done with the cu, i.e. till we call cu__delete(), that now
calls dfops->cu__delete() if it is available.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
There are more things that should be handled differently, such as function
names coming from the .strtab ELF section instead of from the global strings_t
table.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
To cover the case where there aren't structs or even any type at all.
It looks now just like cu__for_each_function, so its more consistent as a bonus
:-)
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Encoding all the non UNDEF OBJECT entries in the symtab. Some must be filtered
in upcoming patches, but for at least kernel/sched.o it works just fine.
To test it I used DaveM's ctfdump and also pdwtags on a --strip-debug, pahole
-Z CTF encoded object.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
For loaders to fill with the address of global variables.
More work is needed to cover relocation, registers, etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
We need it to be able to call cu__for_each_cached_symtab_entry more
than once in the same function.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Finally we can use the Elf file already opened in dwarf_load, call
cu__for_each_cached_symtab_entry to iterate over the symtab entries,
this iterator will first call dwfl_module_getsymtab, that will do the
relocation that will allow us to go from the symtab address to the one
in the DWARF DW_TAG_subprogram tag DW_AT_low_pc attribute.
And voila, for a relatively complex single unit Linux kernel object
file, kernel/sched.o, we go from:
Just DWARF (gcc -g):
$ ls -la kernel/sched.o
1979011 kernel/sched.o
Then we run this to encode the CTF section:
$ pahole -Z kernel/sched.o
And get a file with both DWARF and CTF ELF sections:
$ ls -la kernel/sched.o
2019848 kernel/sched.o
We still need to encode the "OBJECTS", i.e. variables, but this
gets us from 1979011 (just DWARF) to:
$ strip--strip-debug kernel/sched.o
$ ls -la kernel/sched.o
-rw-rw-r-- 1 acme acme 507008 2009-03-30 23:01 kernel/sched.o
25% of the original size.
Of course we don't have inline expansion information, parameter names,
goto labels, etc, but should be good enough for most use cases.
See, without DWARF data, if we ask for it to use DWARF, nothing will be
printed, if we don't speficy the format, it will try first DWARF, it
will not find anything, it will try CTF:
$ pahole -F dwarf kernel/sched.o
$ pahole -C seq_operations kernel/sched.o
struct seq_operations {
void * (*start)(struct seq_file *, loff_t *); /* 0 8 */
void (*stop)(struct seq_file *, void *); /* 8 8 */
void * (*next)(struct seq_file *, void *, loff_t *); /* 16 8 */
int (*show)(struct seq_file *, void *); /* 24 8 */
/* size: 32, cachelines: 1, members: 4 */
/* last cacheline: 32 bytes */
};
$ $ pfunct -Vi -f schedule kernel/sched.o
void schedule(void);
{ /* low_pc=0xe01 */
}/* size: 83 */
$
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
And also export the namespace destructor.
The tag__delete destructor now also uses these new destructors.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
CTF will need this distinction in that it handles functions differently, using
data in the ELF symbol table.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Because we already use ctf__load in libctf.c, rename the others to
disambiguate, and also as there are the __load_dir and __load_files
it looks more consistent.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
So after the next patch, when dwarf_loader will use this new core
functionality, it recognizes:
908 typedef int __m64 __attribute__ ((__vector_size__ (8))); size: 8
909 int array __attribute__ ((__vector_size__ (8))); size: 8
910 int array __attribute__ ((__vector_size__ (4))); size: 4
911 short int array __attribute__ ((__vector_size__ (2))); size: 2
912 char array __attribute__ ((__vector_size__ (1))); size: 1
The above output was obtained using pdwtags.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Nasty trick, but works and should be properly documented in the sources
and here:
If struct namespace.shared_tags is 1, we actually are reusing the list
of enumerators in another namespace, so we shouldn't delete them, for
that list_for_each_tag now means more for each _unshared_ tag, so that
cu__delete doesn't visits it, double freeing enumerator tags.
type__for_each_enumerator knows that and only for enums we'll set this
->shared_tags bit to 1, so we should be safe...
Disgusting? send me a patch, but without increasing memory or processing
footprints, please ;-)
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
CTF doesn't have support for multiple array dimensions, so it flattens
the arrays.
This caused a large number of false positives in ctfdwdiff, so introduce
this conf_fprintf option, use it in pahole and ctfdwdiff.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
For a file with just DWARF info:
$ pahole -F ctf build/pahole
$
But if we ask that it also try dwarf:
$ pahole -F ctf,dwarf build/pahole | head -2
struct _IO_FILE {
int _flags; /* 0 4 */
$
Useful when testing the new CTF support in these tools, as we'll be able to,
from the DWARF info in objects, generate the CTF equivalent and add to the same
object, then run pahole -A -F ctf, pahole -A -F dwarf and compare the outputs.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
So that the user can specify what is the order it wants for decodind, as
we can have several debugging formats encoded in different ELF sections.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Will later be used when generating the CTF info, be it in a separate
file, be it on a new ELF section inserted into this filename.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
To shorten the name and to reflect the fact that we're no longer
"finding" a type, but merely accessing an array with a bounds check in
this function.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
We now create a new integral type (enum or base_types), creating typedef
chains if needed, while caching the bit_size and bit_offset, so that we
can easily reencode the whole file into CTF.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Needed for reencoding DWARF bitfields, where we need to create a new
enum that has a bitfield_size bits size.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This will help us in the next csets when we need to know both the full
size of the base_type used in an bitfield _and_ the size in bits of the
bitfield member.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Because we will need the "bit_offset" and "bit_size" names when converting the
representation of offset and size everywhere to be in bits, not bytes.
At the same time we will keep bitfield_size and bitfield_offset when we convert
from DWARF to CTF and will calculate them when loading CTF, so that the
conversion of the algorithms in dwarves_reorganize, that have all sorts of
subtle issues, can be left for later.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
It makes no sense to try to lookup the abstract_origin (a Dwarf_Off)
after we recode the types just after load.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Amazing how many crept up over time, should have set the
execute bit of .git/hooks/pre-commit already, duh.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
So that we immensely reduce the memory footprint by doing filtering and
other processing/pretty printing as the cus are loaded, discarding them
right away.
The next cset will use this scheme in pahole.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Out of cu__find_struct_by_name so that we can do a string__find
once and lookup the string id on multiple cus.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Also introducing cus__load, that load just one file.
The new cus__load_files routine now iterates thru the provided array
calling cus__load for each, and that in turn will try first dwarf__load,
and if that fail, i.e. if no DWARF info is found, call ctf__load.
This now allows loading DWARF _and_ CTF files at the same time. This
will be useful in the future when we, from DWARF generate CTF and at the
same time do a codiff, comparing the freshly generated CTF file with the
DWARF it came from.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Not used anymore now that cus__loadfl is sanitized. Now we can even
remove the fl (historically comes from libdwfl, when we used to pass an
argp, argh!).
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
So that, when not needing the DWARF info, the apps can tell that at load
time, and then the dwarf loader can just free all the dwarf_tags
allocated, reducing memory usage.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
To visit all parms, lexblocks, namespaces, i.e. not just the top level
tags listed in cu->tags.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
To take advantage of cache effects and to avoid calling cu__find_holes
more than once on the same struct.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
parameter__type was needed because the abstract_origin resolution was
done later, now it is at dwarf recode time, and for debugging formats
that don't have this crap, never. So it now can use the same idiom as
other tags: foo->tag.type.
parameter__name still exists because the tools still want a string
returned, but for some what they want is indeed the string_t, so that
when looking for a particular string it can be done as an string__find
for the key + integer comparision instead of doing a costlier strcmp.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Had to be a big sweeping change, but the regression tests shows just
improvements :-)
Now we stop using an id in struct tag, only storing the type, that now
uses 16 bits only, as CTF does.
Each format loader has to go on adding the types to the core, that
figures out if it is a tag that can be on the tag->type field
(tag__is_tag_type).
Formats that already have the types separated and in sequence, such as
CTF, just ask the core to insert in the types_table directly with its
original ID.
For DWARF, we ask the core to put it on the table, in sequence, and return the
index, that is then stashed with the DWARF specific info (original id, type,
decl_line, etc) and hashed by the original id. Later we recode everything,
looking up via the original type, getting the small_id to put on the tag->type.
The underlying debugging info not needed by the core is stashed in tag->priv,
and the DWARF loader now just allocates sizeof(struct dwarf_tag) at the end of
the core tag and points it there, and makes that info available thru
cu->orig_info. In the future we can ask, when loading a cu, that this info be
trown away, so that we reduce the memory footprint for big multi-cu files such
as the Linux kernel.
There is also a routine to ask for inserting a NULL, as we still have
bugs in the CTF decoding and thus some entries are being lost, to avoid
using an undefined pointer when traversing the types_table the ctf
loader puts a NULL there via cu__table_nullify_type_entry() and then
cu__for_each_type skips those.
There is some more cleanups for leftovers that I avoided cleaning to
reduce this changeset.
And also while doing this I saw that enums can appear without any
enumerators and that an array with DW_TAG_GNU_vector is actually a
different tag, encoded this way till we get to DWARF4 ;-)
So now we don't have to lookup on a hash table looking for DWARF
offsets, we can do the more sensible thing of just indexing the
types_tags array.
Now to do some cleanups and try to get the per cu encoder done. Then
order all the cus per number of type entries, pick the one with more,
then go on merging/recoding the types of the others and putting the
parent linkage in place.
Just to show the extent of the changes:
$ codiff /tmp/libdwarves.so.1.0.0 build/libdwarves.so.1.0.0
/home/acme/git/pahole/dwarves.c:
struct cu | -4048
struct tag | -32
struct ptr_to_member_type | -32
struct namespace | -32
struct type | -32
struct class | -32
struct base_type | -32
struct array_type | -32
struct class_member | -32
struct lexblock | -32
struct ftype | -32
struct function | -64
struct parameter | -32
struct variable | -32
struct inline_expansion | -32
struct label | -32
struct enumerator | -32
17 structs changed
tag__follow_typedef | +3
tag__fprintf_decl_info | +25
array_type__fprintf | +6
type__name | -126
type__find_first_biggest_size_base_type_member | -3
typedef__fprintf | +16
imported_declaration__fprintf | +6
imported_module__fprintf | +3
cu__new | +26
cu__delete | +26
hashtags__hash | -65
hash_64 | -124
hlist_add_head | -78
hashtags__find | -157
cu__hash | -80
cu__add_tag | +20
tag__prefix | -3
cu__find_tag_by_id | -2
cu__find_type_by_id | -3
cu__find_first_typedef_of_type | +38
cu__find_base_type_by_name | +68
cu__find_base_type_by_name_and_size | +72
cu__find_struct_by_name | +59
cus__find_struct_by_name | +8
cus__find_tag_by_id | +5
cus__find_cu_by_name | -6
lexblock__find_tag_by_id | -173
cu__find_variable_by_id | -197
list__find_tag_by_id | -308
cu__find_parameter_by_id | -60
tag__ptr_name | +6
tag__name | +15
variable__type | +13
variable__name | +7
class_member__size | +6
parameter__name | -119
tag__parameter | -14
parameter__type | -143
type__fprintf | -29
union__fprintf | +6
class__add_vtable_entry | -9
type__add_member | -6
type__clone_members | -3
enumeration__add | -6
function__name | -156
ftype__has_parm_of_type | -39
class__find_holes | -27
class__has_hole_ge | -3
type__nr_members_of_type | +3
lexblock__account_inline_expansions | +3
cu__account_inline_expansions | -18
ftype__fprintf_parms | +46
function__tag_fprintf | +24
lexblock__fprintf | -6
ftype__fprintf | +3
function__fprintf_stats | -18
function__size | -6
class__vtable_fprintf | -11
class__fprintf | -21
tag__fprintf | -35
60 functions changed, 513 bytes added, 2054 bytes removed, diff: -1541
/home/acme/git/pahole/ctf_loader.c:
struct ctf_short_type | +0
14 structs changed
type__init | -14
type__new | -9
class__new | -12
create_new_base_type | -7
create_new_base_type_float | -7
create_new_array | -8
create_new_subroutine_type | -9
create_full_members | -18
create_short_members | -18
create_new_class | +1
create_new_union | +1
create_new_enumeration | -19
create_new_forward_decl | -2
create_new_typedef | +3
create_new_tag | -5
load_types | +16
class__fixup_ctf_bitfields | -3
17 functions changed, 21 bytes added, 131 bytes removed, diff: -110
/home/acme/git/pahole/dwarf_loader.c:
17 structs changed
zalloc | -56
tag__init | +3
array_type__new | +20
type__init | -24
class_member__new | +46
inline_expansion__new | +12
class__new | +81
lexblock__init | +19
function__new | +43
die__create_new_array | +20
die__create_new_parameter | +4
die__create_new_label | +4
die__create_new_subroutine_type | +113
die__create_new_enumeration | -21
die__process_class | +79
die__process_namespace | +76
die__create_new_inline_expansion | +4
die__process_function | +147
__die__process_tag | +34
die__process_unit | +56
die__process | +90
21 functions changed, 851 bytes added, 101 bytes removed, diff: +750
/home/acme/git/pahole/dwarves.c:
struct ptr_table | +16
struct cu_orig_info | +32
2 structs changed
tag__decl_line | +68
tag__decl_file | +70
tag__orig_id | +71
ptr_table__init | +46
ptr_table__exit | +37
ptr_table__add | +183
ptr_table__add_with_id | +165
ptr_table__entry | +64
cu__table_add_tag | +171
cu__table_nullify_type_entry | +38
10 functions changed, 913 bytes added, diff: +913
/home/acme/git/pahole/ctf_loader.c:
2 structs changed
tag__alloc | +52
1 function changed, 52 bytes added, diff: +52
/home/acme/git/pahole/dwarf_loader.c:
struct dwarf_tag | +48
struct dwarf_cu | +4104
4 structs changed
dwarf_cu__init | +83
hashtags__hash | +61
hash_64 | +124
hlist_add_head | +78
hashtags__find | +161
cu__hash | +95
tag__is_tag_type | +171
tag__is_type | +85
tag__is_union | +28
tag__is_struct | +57
tag__is_typedef | +28
tag__is_enumeration | +28
dwarf_cu__find_tag_by_id | +56
dwarf_cu__find_type_by_id | +63
tag__alloc | +114
__tag__print_type_not_found | +108
namespace__recode_dwarf_types | +346
tag__namespace | +14
tag__has_namespace | +86
tag__is_namespace | +28
type__recode_dwarf_specification | +182
tag__type | +14
__tag__print_abstract_origin_not_found | +105
ftype__recode_dwarf_types | +322
tag__ftype | +14
tag__parameter | +14
lexblock__recode_dwarf_types | +736
tag__lexblock | +14
tag__label | +14
tag__recode_dwarf_type | +766
tag__ptr_to_member_type | +14
cu__recode_dwarf_types_table | +88
cu__recode_dwarf_types | +48
dwarf_tag__decl_file | +77
strings__ptr | +33
dwarf_tag__decl_line | +59
dwarf_tag__orig_id | +59
dwarf_tag__orig_type | +59
38 functions changed, 4432 bytes added, diff: +4432
build/libdwarves.so.1.0.0:
147 functions changed, 6782 bytes added, 2286 bytes removed, diff: +4496
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Do nothing for now, just to reduce the size of the upcoming
type recoding patch, aka dwarves undwarvification.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Grrr, the previous commit has the other bits, and as I already pushed it
out publicly... <BROWN PAPER BAG ALERT!> here goes the rests. So much
for bissectability. Sigh.
But the regression test showed only one problem, in C++ code, that I'll
fix in the next commits.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
So that we can then decide in what hashtable we will add it, and this
also paves the way for a type array that will help us in reducing the
size of struct tag by removing the id field.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
And comment the difference to tag__is_type:
tag__is_type == is this tag derived from the 'type' class?
tag__is_tag_type == is this tag a possible type for a tag, i.e.
one we will find in struct tag->type?
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Needed for CTF, where we can have many base types with name "unsigned",
but with different bit sizes, to implement bitfields.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Now we just pass a NULL terminated array of filenames, since we got rid
of that ugly -e insertion hack.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
And make the dwarves use it, so that we can remove duplicate strings in
a multi-CU file (vmlinux anyone?) and have it ready for insertion in a
compressed DWARF format with just the types, or better, CTF or some new
compressed debugging info format.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
So that one can get an skeleton from where a function can be
reimplemented, or a probe can be written to attach to a tracepoint.
Right now it will only expand the types for
struct/union/typedef/enumeration types, but it is a good start.
[acme@doppio pahole]$ pfunct --expand_types --function inet6_ioctl ipv6.ko > a.c
[acme@doppio pahole]$ echo "int main(void) { return 0; }" >> a.c
[acme@doppio pahole]$ gcc -Wall -g a.c -o a
[acme@doppio pahole]$ grep ^#include a.c
[acme@doppio pahole]$
No errors, no includes.
This is present in ctracer, where we don't want to _require_ any header
files, just the object file with the function we want to probe. From
there we get the function signature, and reconstruct the types needed to
access members of structs passed as parameters.
We still need to add padding to reconstruct __attribute__ alignment
effects.
Also, if we can detect what are the exact members accessed in the probe,
we can reconstruct just what is needed to access those members,
hopefully reducing the time needed for gcc to digest the resulting
source code. And also reducing the size of the output, which can
hopefully be interesting to help focus on what the probe is doing.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Basically a wrapper for ftype__fprintf(&function__proto, ...) for the
cases we want the prototype rendered to a buffer, not to a file.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
For correctly created and completely parsed debugging information the type will
always be found, but as we still need to parse more tags and expecting
debugging information to be always correctly built is not sane... sprinkle some
asserts.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This is trying to get CTF friendly, where bitfields are not stored in the
equivalent to the DW_TAG_member dwarf TAG, but on "base types" with bit sizes
different than the real in the DWARF sense, base types (char, long, etc).
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
In libdwarves.so well continue using DW_TAG_ entries and types for now, but its
becoming non-DWARF specific as will be demonstrated with the introduction of
ctf_loader.c in the upcoming csets.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Almost halves the time spent on processing a x86_64 vmlinux. Good, we
have features, now lets have performance ;-)
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
What a mouthful ;-) To be used in finding the most aligned member in a non-packed
type, i.e. one that originally wasn't __attribute__((packed)).
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Arnaldo Carvalho de Melo