GDB has a workaround for DWARF output by ICC, related to missing
DW_AT_declaration on incomplete types. The bug was fixed in ICC 14,
so this commit adjusts GDB accordingly.
For the version check, this adds a new parser function for the ICC
producer string. While at it, it also adds unit tests for the
producer parsing covering the new function and preexisting parsers.
gdb/ChangeLog:
2017-09-26 Walfred Tedeschi <walfred.tedeschi@intel.com>
Pedro Alves <palves@redhat.com>
* dwarf2read.c (dwarf2_cu): Remove field producer_is_icc and add
producer_is_icc_lt_14.
(producer_is_icc_lt_14): New function.
(check_producer): Add code for checking version of ICC.
(producer_is_icc): Move to producer.c.
(read_structure_type): Restrict ICC workaround to ICC<14.
* producer.c: Include selftest.h.
(producer_is_icc, producer_parsing_tests, _initialize_producer):
New functions.
* producer.h (producer_is_icc): New declaration.
copy_string does the exact same thing as savestring, so replace the
usages of the former with the latter.
gdb/ChangeLog:
* dwarf2read.c (copy_string): Remove.
(parse_macro_definition): Replace copy_string with savestring.
This changes add_using_directive to accept a std::vector and then
changes the callers. This allows removing a cleanup.
ChangeLog
2017-09-09 Tom Tromey <tom@tromey.com>
* namespace.h (add_using_directive): Update.
* namespace.c (add_using_directive): Change type of excludes to
std::vector.
* dwarf2read.c (read_import_statement): Use std::vector.
(read_namespace): Update.
* cp-namespace.c (cp_scan_for_anonymous_namespaces): Update.
Since at least 7.3 the "fnfields" field in struct field_info has been
unused. This patch simply removes it.
gdb/ChangeLog:
* dwarf2read.c (struct field_info) <fnfields>: Remove unused
field.
The DWARF reader is littered with the following idiom to read a linkage name
from the debug info:
mangled = dwarf2_string_attr (die, DW_AT_linkage_name, cu);
if (mangled == NULL)
mangled = dwarf2_string_attr (die, DW_AT_MIPS_linkage_name, cu);
This patch introduces functions to simplify this to:
mangled = dw2_linkage_name (die, cu);
or
attr = dw2_linkage_name_attr (die, cu);
gdb/ChangeLog:
* dwarf2read.c (dw2_linkage_name_attr): New function.
(dw2_linkage_name): New function.
(dwarf2_compute_name, dwarf2_physname, read_call_site_scope)
(guess_full_die_structure_name, dwarf2_name): Use dw2_linkage_name.
(anonymous_struct_prefix, dwarf2_name): Use dw2_linkage_name_attr.
Trying to print a function local static variable of a const-qualified
method still doesn't work after the previous fixes:
(gdb) p 'S::method() const'::static_var
$1 = {i1 = 1, i2 = 2, i3 = 3}
(gdb) p S::method() const::static_var
No symbol "static_var" in specified context.
The reason is that the expression parser/evaluator loses the "const",
and the above unquoted case is just like trying to print a variable of
the non-const overload, if it exists, even. As if the above unquoted
case had been written as:
(gdb) p S::method()::static_var
No symbol "static_var" in specified context.
We can see the problem without static vars in the picture. With:
struct S
{
void method ();
void method () const;
};
Compare:
(gdb) print 'S::method(void) const'
$1 = {void (const S * const)} 0x400606 <S::method() const>
(gdb) print S::method(void) const
$2 = {void (S * const)} 0x4005d8 <S::method()> # wrong method!
That's what we need to fix. If we fix that, the function local static
case starts working.
The grammar production for function/method types is this one:
exp: exp '(' parameter_typelist ')' const_or_volatile
This results in a TYPE_INSTANCE expression evaluator operator. For
the example above, we get something like this ("set debug expression 1"):
...
0 TYPE_INSTANCE 1 TypeInstance: Type @0x560fda958be0 (void)
5 OP_SCOPE Type @0x560fdaa544d8 (S) Field name: `method'
...
While evaluating TYPE_INSTANCE, we end up in
value_struct_elt_for_reference, trying to find the method named
"method" that has the prototype recorded in TYPE_INSTANCE. In this
case, TYPE_INSTANCE says that we're looking for a method that has
"(void)" as parameters (that's what "1 TypeInstance: Type
@0x560fda958be0 (void)" above means. The trouble is that nowhere in
this mechanism do we communicate to value_struct_elt_for_reference
that we're looking for the _const_ overload.
value_struct_elt_for_reference only compared parameters, and the
non-const "method()" overload has matching parameters, so it's
considered the right match...
Conveniently, the "const_or_volatile" production in the grammar
already records "const" and "volatile" info in the type stack. The
type stack is not used in this code path, but we can borrow the
information. The patch converts the info in the type stack to an
"instance flags" enum, and adds that as another element in
TYPE_INSTANCE operators. This type instance flags is then applied to
the temporary type that is passed to value_struct_elt_for_reference
for matching.
The other side of the problem is that methods in the debug info aren't
marked const/volatile, so with that in place, the matching never finds
const/volatile-qualified methods.
The problem is that in the DWARF, there's no indication at all whether
a method is const/volatile qualified... For example (c++filt applied
to the linkage name for convenience):
<2><d3>: Abbrev Number: 6 (DW_TAG_subprogram)
<d4> DW_AT_external : 1
<d4> DW_AT_name : (indirect string, offset: 0x3df): method
<d8> DW_AT_decl_file : 1
<d9> DW_AT_decl_line : 58
<da> DW_AT_linkage_name: (indirect string, offset: 0x5b2): S::method() const
<de> DW_AT_declaration : 1
<de> DW_AT_object_pointer: <0xe6>
<e2> DW_AT_sibling : <0xec>
I see the same with both GCC and Clang. The patch works around this
by extracting the cv qualification from the "const" and "volatile" in
the demangled name. This will need further tweaking for "&" and
"const &" overloads, but we don't support them in the parser yet,
anyway.
The TYPE_CONST changes were necessary otherwise the comparisons in valops.c:
if (TYPE_CONST (intype) != TYPE_FN_FIELD_CONST (f, j))
continue;
would fail, because when both TYPE_CONST() TYPE_FN_FIELD_CONST() were
true, their values were different.
BTW, I'm recording the const/volatile-ness of methods in the
TYPE_FN_FIELD info because #1 - I'm not sure it's kosher to change the
method's type directly (vs having to call make_cv_type to create a new
type), and #2 it's what stabsread.c does:
...
case 'A': /* Normal functions. */
new_sublist->fn_field.is_const = 0;
new_sublist->fn_field.is_volatile = 0;
(*pp)++;
break;
case 'B': /* `const' member functions. */
new_sublist->fn_field.is_const = 1;
new_sublist->fn_field.is_volatile = 0;
...
After all this, this finally all works:
print S::method(void) const
$1 = {void (const S * const)} 0x400606 <S::method() const>
(gdb) p S::method() const::static_var
$2 = {i1 = 1, i2 = 2, i3 = 3}
gdb/ChangeLog:
2017-09-04 Pedro Alves <palves@redhat.com>
* c-exp.y (function_method, function_method_void): Add current
instance flags to TYPE_INSTANCE.
* dwarf2read.c (check_modifier): New.
(compute_delayed_physnames): Assert that only C++ adds delayed
physnames. Mark fn_fields as const/volatile depending on
physname.
* eval.c (make_params): New type_instance_flags parameter. Use
it as the new type's instance flags.
(evaluate_subexp_standard) <TYPE_INSTANCE>: Extract the instance
flags element and pass it to make_params.
* expprint.c (print_subexp_standard) <TYPE_INSTANCE>: Handle
instance flags element.
(dump_subexp_body_standard) <TYPE_INSTANCE>: Likewise.
* gdbtypes.h: Include "enum-flags.h".
(type_instance_flags): New enum-flags type.
(TYPE_CONST, TYPE_VOLATILE, TYPE_RESTRICT, TYPE_ATOMIC)
(TYPE_CODE_SPACE, TYPE_DATA_SPACE): Return boolean.
* parse.c (operator_length_standard) <TYPE_INSTANCE>: Adjust.
(follow_type_instance_flags): New function.
(operator_check_standard) <TYPE_INSTANCE>: Adjust.
* parser-defs.h (follow_type_instance_flags): Declare.
* valops.c (value_struct_elt_for_reference): const/volatile must
match too.
gdb/testsuite/ChangeLog:
2017-09-04 Pedro Alves <palves@redhat.com>
* gdb.base/func-static.c (S::method const, S::method volatile)
(S::method volatile const): New methods.
(c_s, v_s, cv_s): New instances.
(main): Call method() on them.
* gdb.base/func-static.exp (syntax_re, cannot_resolve_re): New variables.
(cannot_resolve): New procedure.
(cxx_scopes_list): Test cv methods. Add print-scope-quote and
print-quote-unquoted columns.
(do_test): Test printing each scope too.
An earlier commit made GDB no longer assume no-debug-info functions
return int. This commit gives the same treatment to variables.
Currently, you can end misled by GDB over output like this:
(gdb) p var
$1 = -1
(gdb) p /x var
$2 = 0xffffffff
until you realize that GDB is assuming that the variable is an "int",
because:
(gdb) ptype var
type = <data variable, no debug info>
You may try to fix it by casting, but that doesn't really help:
(gdb) p /x (unsigned long long) var
$3 = 0xffffffffffffffff # incorrect
^^
That's incorrect output, because the variable was defined like this:
uint64_t var = 0x7fffffffffffffff;
^^
What happened is that with the cast, GDB did an int -> 'unsigned long
long' conversion instead of reinterpreting the variable as the cast-to
type. To get at the variable properly you have to reinterpret the
variable's address manually instead, with either:
(gdb) p /x *(unsigned long long *) &var
$4 = 0x7fffffffffffffff
(gdb) p /x {unsigned long long} &var
$5 = 0x7fffffffffffffff
After this commit GDB does it for you. This is what you'll get
instead:
(gdb) p var
'var' has unknown type; cast it to its declared type
(gdb) p /x (unsigned long long) var
$1 = 0x7fffffffffffffff
As in the functions patch, the "compile" machinery doesn't currently
have the cast-to type handy, so it continues assuming no-debug
variables have int type, though now at least it warns.
The change to gdb.cp/m-static.exp deserves an explanation:
- gdb_test "print 'gnu_obj_1::method()::sintvar'" "\\$\[0-9\]+ = 4" \
+ gdb_test "print (int) 'gnu_obj_1::method()::sintvar'" "\\$\[0-9\]+ = 4" \
That's printing the "sintvar" function local static of the
"gnu_obj_1::method()" method.
The problem with that test is that that "'S::method()::static_var'"
syntax doesn't really work in C++ as you'd expect. The way to make it
work correctly currently is to quote the method part, not the whole
expression, like:
(gdb) print 'gnu_obj_1::method()'::sintvar
If you wrap the whole expression in quotes, like in m-static.exp, what
really happens is that the parser considers the whole string as a
symbol name, but there's no debug symbol with that name. However,
local statics have linkage and are given a mangled name that demangles
to the same string as the full expression, so that's what GDB prints.
After this commit, and without the cast, the print in m-static.exp
would error out saying that the variable has unknown type:
(gdb) p 'gnu_obj_1::method()::sintvar'
'gnu_obj_1::method()::sintvar' has unknown type; cast it to its declared type
TBC, if currently (even before this series) you try to print any
function local static variable of type other than int, you'll get
bogus results. You can see that with m-static.cc as is, even.
Printing the "svar" local, which is a boolean (1 byte) still prints as
"int" (4 bytes):
(gdb) p 'gnu_obj_1::method()::svar'
$1 = 1
(gdb) ptype 'gnu_obj_1::method()::svar'
type = <data variable, no debug info>
This probably prints some random bogus value on big endian machines.
If 'svar' was of some aggregate type (etc.) we'd still print it as
int, so the problem would have been more obvious... After this
commit, you'll get instead:
(gdb) p 'gnu_obj_1::method()::svar'
'gnu_obj_1::method()::svar' has unknown type; cast it to its declared type
... so at least GDB is no longer misleading. Making GDB find the real
local static debug symbol is the subject of the following patches. In
the end, it'll all "Just Work".
gdb/ChangeLog:
2017-09-04 Pedro Alves <palves@redhat.com>
* ax-gdb.c: Include "typeprint.h".
(gen_expr_for_cast): New function.
(gen_expr) <OP_CAST, OP_CAST_TYPE>: Use it.
<OP_VAR_VALUE, OP_MSYM_VAR_VALUE>: Error out if the variable's
type is unknown.
* dwarf2read.c (new_symbol_full): Fallback to int instead of
nodebug_data_symbol.
* eval.c: Include "typeprint.h".
(evaluate_subexp_standard) <OP_VAR_VALUE, OP_VAR_MSYM_VALUE>:
Error out if symbol has unknown type.
<UNOP_CAST, UNOP_CAST_TYPE>: Common bits factored out to
evaluate_subexp_for_cast.
(evaluate_subexp_for_address, evaluate_subexp_for_sizeof): Handle
OP_VAR_MSYM_VALUE.
(evaluate_subexp_for_cast): New function.
* gdbtypes.c (init_nodebug_var_type): New function.
(objfile_type): Use it to initialize types of variables with no
debug info.
* typeprint.c (error_unknown_type): New.
* typeprint.h (error_unknown_type): New declaration.
* compile/compile-c-types.c (convert_type_basic): Handle
TYPE_CODE_ERROR; warn and fallback to int for variables with
unknown type.
gdb/testsuite/ChangeLog:
2017-09-04 Pedro Alves <palves@redhat.com>
* gdb.asm/asm-source.exp: Add casts to int.
* gdb.base/nodebug.c (dataglobal8, dataglobal32_1, dataglobal32_2)
(dataglobal64_1, dataglobal64_2): New globals.
* gdb.base/nodebug.exp: Test different expressions involving the
new globals, with print, whatis and ptype. Add casts to int.
* gdb.base/solib-display.exp: Add casts to int.
* gdb.compile/compile-ifunc.exp: Expect warning. Add cast to int.
* gdb.cp/m-static.exp: Add cast to int.
* gdb.dwarf2/dw2-skip-prologue.exp: Add cast to int.
* gdb.threads/tls-nodebug.exp: Check that gdb errors out printing
tls variable with no debug info without a cast. Test with a cast
to int too.
* gdb.trace/entry-values.exp: Add casts.
GDB was now accessing as signatured_type memory allocated only by size of
dwarf2_per_cu_data.
gdb/ChangeLog
2017-08-24 Jan Kratochvil <jan.kratochvil@redhat.com>
* dwarf2read.c (build_type_psymtabs_reader): New prototype.
(process_psymtab_comp_unit): Accept IS_DEBUG_TYPES.
(read_comp_units_from_section): New parameter abbrev_section, use
read_and_check_comp_unit_head, allocate signatured_type if needed.
(create_all_comp_units): Update read_comp_units_from_section caller.
This changes gdb_realpath to return a unique_xmalloc_ptr and fixes up
the callers. This allows removing some cleanups. This change by
itself caused xfullpath.exp to fail; and attempting to fix that ran
into various problems (like .get() being optimized out); so this patch
also rewrites xfullpath.exp to be a C++ selftest instead.
ChangeLog
2017-08-22 Tom Tromey <tom@tromey.com>
* exec.c (exec_file_attach): Update.
* linux-thread-db.c (try_thread_db_load): Update.
* guile/scm-safe-call.c (gdbscm_safe_source_script): Update.
* utils.c (gdb_realpath): Change return type.
(gdb_realpath_keepfile): Update.
(gdb_realpath_check_trailer, gdb_realpath_tests): New functions.
(_initialize_utils): Register the new self test.
* source.c (openp): Update.
(find_and_open_source): Update.
* nto-tdep.c (nto_find_and_open_solib): Update.
* main.c (set_gdb_data_directory): Update.
(captured_main_1): Update.
* dwarf2read.c (dwarf2_get_dwz_file): Update
(dw2_map_symbol_filenames): Update.
* auto-load.c (auto_load_safe_path_vec_update): Update.
(filename_is_in_auto_load_safe_path_vec): Change type of
"filename_realp".
(auto_load_objfile_script): Update.
(file_is_auto_load_safe): Update. Use std::string.
* utils.h (gdb_realpath): Return a gdb::unique_xmalloc_ptr.
testsuite/ChangeLog
2017-08-22 Tom Tromey <tom@tromey.com>
* gdb.gdb/xfullpath.exp: Remove.
This plugs a couple leaks introduced by commit fff8551cf5
("dwarf2read.c: Some C++fycation, use std::vector, std::unique_ptr").
The first problem is that nothing owns the temporary line_header that
handle_DW_AT_stmt_list creates in some cases. Before the commit
mentioned above, the temporary line_header case used to have:
make_cleanup (free_cu_line_header, cu);
and that cleanup was assumed to be run by process_die, after
handle_DW_AT_stmt_list returns and before child DIEs were processed.
The second problem is found in setup_type_unit_groups: that also used
to have a similar make_cleanup call, and ended up with a similar leak
after the commit mentioned above.
Fix both cases by recording in dwarf2_cu whether a line header is
owned by the cu/die, and have process_die explicitly free the
line_header if so, making use of a new RAII object that also replaces
the reset_die_in_process cleanup, while at it.
Thanks to Philippe Waroquiers for noticing the leak and pointing in
the right direction.
gdb/ChangeLog:
2017-08-17 Pedro Alves <palves@redhat.com>
* dwarf2read.c (struct dwarf2_cu) <line_header_die_owner>: New
field.
(reset_die_in_process): Delete, replaced by ...
(process_die_scope): ... this new class. Make it responsible for
freeing cu->line_header too.
(process_die): Use process_die_scope.
(handle_DW_AT_stmt_list): Record the line header's owner CU/DIE in
cu->line_header_die_owner. Don't release the line header if it's
owned by the CU.
(setup_type_unit_groups): Make the CU/DIE own the line header.
Don't release the line header here.
The dwarf2_string_attr did not allow DW_FORM_GNU_str_index as a form for
string types. This manifested as null strings in the namespace_name
lookup (replaced with "(anonymous namespace)") when debugging
Fission-compiled code.
gdb/ChangeLog:
* dwarf2read.c (dwarf2_string_attr): Allow DW_FORM_GNU_strp_alt.
gdb/ChangeLog:
2017-07-20 Pedro Alves <palves@redhat.com>
* dwarf2read.c (dw2_lookup_symbol): Use
SYMBOL_MATCHES_SEARCH_NAME.
* psymtab.c (psym_lookup_symbol): Use SYMBOL_MATCHES_SEARCH_NAME.
Turns out somewhere along the refactoring of the multiple-CU support
for Fission I broke the patch before submitting it (& seems to have
broken Fission support generally).
Syncing back to the point at which the patch was committed, the
previous test results on my machine are:
expected passes: 36137
unexpected failures: 416
with the previous (broken) patch committed:
expected passes: 36131
unexpected failures: 429
With this one line patch applied on top of the broken commit:
expected passes: 36144
unexpected failures: 416
(& all other result counts remained the same in all 3 cases)
gdb/ChangeLog:
2017-07-18 David Blaikie <dblaikie@gmail.com>
* dwarf2read.c (create_cus_hash_table): Re-add lost initialization
of dwo_cu's dwo_file.
Tab completion when debugging a program binary that uses GDB index is
surprisingly much slower than when GDB uses psymtabs instead. Around
1.5x/3x slower. That's surprising, because the whole point of GDB
index is to speed things up...
For example, with:
set pagination off
set $count = 0
while $count < 400
complete b string_prin # matches gdb's string_printf
printf "count = %d\n", $count
set $count = $count + 1
end
$ time ./gdb --batch -q ./gdb-with-index -ex "source script.cmd"
real 0m11.042s
user 0m10.920s
sys 0m0.042s
$ time ./gdb --batch -q ./gdb-without-index -ex "source script.cmd"
real 0m4.635s
user 0m4.590s
sys 0m0.037s
Same but with:
- complete b string_prin
+ complete b zzzzzz
to exercise the no-matches worst case, master currently gets you
something like:
with index without index
real 0m11.971s 0m8.413s
user 0m11.912s 0m8.355s
sys 0m0.035s 0m0.035s
Running gdb under perf shows 80% spent inside
maybe_add_partial_symtab_filename, and 20% spent in the lbasename
inside that.
The problem that tab completion walks over all compunit symtabs, and
for each, walks the contained file symtabs. And there a huge number
of file symtabs (each included system header, etc.) that appear in
each compunit symtab's file symtab list. As in, when debugging GDB, I
have 367381 symtabs iterated, when of those only 5371 filenames are
unique...
This was a regression from the earlier (nice) split of symtabs in
compunit symtabs + file symtabs.
The fix here is to add a cache of unique filenames per objfile so that
the walk / uniquing is only done once. There's already a abstraction
for this in symtab.c; this patch moves that code out to a separate
file and C++ifies it bit.
This makes the worst-case scenario above consistently drop to ~2.5s
(1.5s for the "string_prin" hit case), making it over 3.3x times
faster than psymtabs in this use case (7x in the "string_prin" hit
case).
gdb/ChangeLog:
2017-07-17 Pedro Alves <palves@redhat.com>
* Makefile.in (COMMON_OBS): Add filename-seen-cache.o.
* dwarf2read.c: Include "filename-seen-cache.h".
* dwarf2read.c (dwarf2_per_objfile) <filenames_cache>: New field.
(dw2_map_symbol_filenames): Build and use a filenames_seen_cache.
* filename-seen-cache.c: New file.
* filename-seen-cache.h: New file.
* symtab.c: Include "filename-seen-cache.h".
(struct filename_seen_cache, INITIAL_FILENAME_SEEN_CACHE_SIZE)
(create_filename_seen_cache, clear_filename_seen_cache)
(delete_filename_seen_cache, filename_seen): Delete, parts moved
to filename-seen-cache.h/filename-seen-cache.c.
(output_source_filename, sources_info)
(maybe_add_partial_symtab_filename)
(make_source_files_completion_list): Adjust to use
filename_seen_cache.
This makes dwarf2_per_objfile a class with cdtors.
A following patch will add a non-trivial field to struct
dwarf2_per_objfile, making dwarf2_per_objfile itself non-trivial.
Since dwarf2_per_objfile is allocated in an obstack, we need to run
its cdtors manually.
Tested on x86-64 GNU/Linux.
gdb/ChangeLog:
2017-07-17 Pedro Alves <palves@redhat.com>
* dwarf2read.c (dwarf2_per_objfile): In-class initialize all
fields.
(dwarf2_per_objfile::dwarf2_per_objfile(objfile*, const
dwarf2_debug_sections*)): New.
(dwarf2_per_objfile::dwarf2_per_objfile(const
dwarf2_per_objfile&)): Declare as deleted.
(dwarf2_per_objfile::operator=): Declare as deleted.
(dwarf2_per_objfile::dwarf2_per_objfile)
(dwarf2_per_objfile::~dwarf2_per_objfile)
(dwarf2_per_objfile::free_cached_comp_units): New.
(dwarf2_has_info): dwarf2_per_objfile initialization code moved to
ctor. Call dwarf2_per_objfile's ctor manually.
(dwarf2_locate_sections): Deleted/refactored as ...
(dwarf2_per_objfile::locate_sections): ... this new method.
(free_cached_comp_units): Defer to
dwarf2_per_objfile::free_cached_comp_units.
(dwarf2_free_objfile): Call dwarf2_per_objfile's dtor manually.
In some cases a compiler may produce a single object file (& thus single
DWO file) representing multiple source files. The most common example of
this is in whole program optimization (such as LLVM's LTO). Fission may
still be a beneficial feature to use here - to avoid the need to
read/link the debug info with system libraries and the like.
This change adds basic support for multiple CUs in a single DWO file to
support LLVM's output in this situation.
There is still outstanding work to design and implement a solution for
cross-CU references (usually using DW_FORM_ref_addr) in this scenario.
For now LLVM works around this by duplicating DIEs rather than making
cross-CU references in DWO files. This degrades debugger
behavior/quality especially for file-local entities.
2017-07-06 David Blaikie <dblaikie@gmail.com>
* dwarf2read.c (struct dwo_file): Use a htab of dwo_unit* (rather than
a singular dwo_unit*) to support multiple CUs in the same way that
multiple TUs are supported.
(create_cus_hash_table): Replace create_dwo_cu with a function for
parsing multiple CUs from a DWO file.
(open_and_init_dwo_file): Use create_cus_hash_table rather than
create_dwo_cu.
(lookup_dwo_cutu): Lookup CU in the hash table in the dwo_file with
htab_find, rather than comparing the signature to a singleton CU in
the dwo_file.
2017-07-06 David Blaikie <dblaikie@gmail.com>
* gdb.dwarf2/fission-multi-cu.S: Test containing multiple CUs in a DWO,
built from fissiont-multi-cu{1,2}.c.
* gdb.dwarf2/fission-multi-cu.exp: Test similar to fission-base.exp,
except putting 'main' and 'func' in separate CUs in the same DWO file.
* gdb.dwarf2/fission-multi-cu1.c: First CU for the multi-CU-single-DWO
test.
* gdb.dwarf2/fission-multi-cu2.c: Second CU in the multi-CU-single-DWO
test.
This commit eliminates make_cleanup_obstack_free, replacing it with a
new auto_obstack type that inherits obstack to add cdtors.
These changes in the parsers may not be obvious:
- obstack_init (&name_obstack);
- make_cleanup_obstack_free (&name_obstack);
+ name_obstack.clear ();
Here, the 'name_obstack' variable is a global. The change means that
the obstack's contents from a previous parse will stay around until
the next parsing starts. I.e., memory won't be reclaimed until then.
I don't think that's a problem, these objects don't really grow much
at all.
The other option I tried was to add a separate type that is like
auto_obstack but manages an external obstack, just for those cases. I
like the current approach better as that other approach adds more
boilerplate and yet another type to learn.
gdb/ChangeLog:
2017-06-27 Pedro Alves <palves@redhat.com>
* c-exp.y (name_obstack): Now an auto_obstack.
(yylex): Use auto_obstack::clear.
(c_parse): Use auto_obstack::clear instead of reinitializing and
freeing the obstack.
* c-lang.c (evaluate_subexp_c): Use auto_obstack.
* d-exp.y (name_obstack): Now an auto_obstack.
(yylex): Use auto_obstack::clear.
(d_parse): Use auto_obstack::clear instead of reinitializing and
freeing the obstack.
* dwarf2loc.c (fetch_const_value_from_synthetic_pointer): Use
auto_obstack.
* dwarf2read.c (create_addrmap_from_index)
(dwarf2_build_psymtabs_hard)
(update_enumeration_type_from_children): Likewise.
* gdb_obstack.h (auto_obstack): New type.
* go-exp.y (name_obstack): Now an auto_obstack.
(build_packaged_name): Use auto_obstack::clear.
(go_parse): Use auto_obstack::clear instead of reinitializing and
freeing the obstack.
* linux-tdep.c (linux_make_mappings_corefile_notes): Use
auto_obstack.
* printcmd.c (printf_wide_c_string, ui_printf): Use auto_obstack.
* rust-exp.y (work_obstack): Now an auto_obstack.
(rust_parse, rust_lex_tests): Use auto_obstack::clear instead of
reinitializing and freeing the obstack.
* utils.c (do_obstack_free, make_cleanup_obstack_free): Delete.
(host_char_to_target): Use auto_obstack.
* utils.h (make_cleanup_obstack_free): Delete declaration.
* valprint.c (generic_emit_char, generic_printstr): Use
auto_obstack.
We should close the file before unlinking because on MS-Windows one
cannot delete a file that is still open.
I considered making 'gdb::unlinker::unlinker(const char *)'
'noexcept(true)' and then adding
static_assert (noexcept (gdb::unlinker (filename.c_str ())), "");
but that doesn't really work because gdb::unlinker has a gdb_assert,
which can throw a QUIT if/when the assertion fails. 'noexcept(true)'
would cause GDB to abruptly terminate if/when the assertion fails.
gdb/ChangeLog:
2017-06-19 Pedro Alves <palves@redhat.com>
* dwarf2read.c (write_psymtabs_to_index): Construct file_closer
after gdb::unlinker.
In some cases we've been replacing heap-allocated gdb_byte buffers
managed with xmalloc/make_cleanup(xfree) with gdb::vector<gdb_byte>.
That usually pessimizes the code a little bit because std::vector
value-initializes elements (which for gdb_byte means
zero-initialization), while if you're creating a temporary buffer,
you're most certaintly going to fill it in with some data. An
alternative is to use
unique_ptr<gdb_byte[]> buf (new gdb_byte[size]);
but it looks like that's not very popular.
Recently, a use of obstacks in dwarf2read.c was replaced with
std::vector<gdb_byte> and that as well introduced a pessimization for
always memsetting the buffer when it's garanteed that the zeros will
be overwritten immediately. (see dwarf2read.c change in this patch to
find it.)
So here's a different take at addressing this issue "by design":
#1 - Introduce default_init_allocator<T>
I.e., a custom allocator that does default construction using default
initialization, meaning, no more zero initialization. That's the
default_init_allocation<T> class added in this patch.
See "Notes" at
<http://en.cppreference.com/w/cpp/container/vector/resize>.
#2 - Introduce def_vector<T>
I.e., a convenience typedef, because typing the allocator is annoying:
using def_vector<T> = std::vector<T, gdb::default_init_allocator<T>>;
#3 - Introduce byte_vector
Because gdb_byte vectors will be the common thing, add a convenience
"byte_vector" typedef:
using byte_vector = def_vector<gdb_byte>;
which is really the same as:
std::vector<gdb_byte, gdb::default_init_allocator<gdb_byte>>;
The intent then is to make "gdb::byte_vector" be the go-to for dynamic
byte buffers. So the less friction, the better.
#4 - Adjust current code to use it.
To set the example going forward. Replace std::vector uses and also
unique_ptr<byte[]> uses.
One nice thing is that with this allocator, for changes like these:
-std::unique_ptr<byte[]> buf (new gdb_byte[some_size]);
+gdb::byte_vector buf (some_size);
fill_with_data (buf.data (), buf.size ());
the generated code is the same as before. I.e., the compiler
de-structures the vector and gets rid of the unused "reserved vs size"
related fields.
The other nice thing is that it's easier to write
gdb::byte_vector buf (size);
than
std::unique_ptr<gdb_byte[]> buf (new gdb_byte[size]);
or even (C++14):
auto buf = std::make_unique<gdb_byte[]> (size); // zero-initializes...
#5 - Suggest s/std::vector<gdb_byte>/gdb::byte_vector/ going forward.
Note that this commit actually fixes a couple of bugs where the current
code is incorrectly using "std::vector::reserve(new_size)" and then
accessing the vector's internal buffer beyond the vector's size: see
dwarf2loc.c and charset.c. That's undefined behavior and may trigger
debug mode assertion failures. With default_init_allocator,
"resize()" behaves like "reserve()" performance wise, in that it
leaves new elements with unspecified values, but, it does that safely
without triggering undefined behavior when you access those values.
gdb/ChangeLog:
2017-06-14 Pedro Alves <palves@redhat.com>
* ada-lang.c: Include "common/byte-vector.h".
(ada_value_primitive_packed_val): Use gdb::byte_vector.
* charset.c (wchar_iterator::iterate): Resize the vector instead
of reserving it.
* common/byte-vector.h: Include "common/def-vector.h".
(wchar_iterator::m_out): Now a gdb::def_vector<gdb_wchar_t>.
* cli/cli-dump.c: Include "common/byte-vector.h".
(dump_memory_to_file, restore_binary_file): Use gdb::byte_vector.
* common/byte-vector.h: New file.
* common/def-vector.h: New file.
* common/default-init-alloc.h: New file.
* dwarf2loc.c: Include "common/byte-vector.h".
(rw_pieced_value): Use gdb::byte_vector, and resize the vector
instead of reserving it.
* dwarf2read.c: Include "common/byte-vector.h".
(data_buf::m_vec): Now a gdb::byte_vector.
* gdb_regex.c: Include "common/def-vector.h".
(compiled_regex::compiled_regex): Use gdb::def_vector<char>.
* mi/mi-main.c: Include "common/byte-vector.h".
(mi_cmd_data_read_memory): Use gdb::byte_vector.
* printcmd.c: Include "common/byte-vector.h".
(print_scalar_formatted): Use gdb::byte_vector.
* valprint.c: Include "common/byte-vector.h".
(maybe_negate_by_bytes, print_decimal_chars): Use
gdb::byte_vector.
... instead of vector of pointers
There's no real reason for having mapped_symtab::data be a vector of
heap-allocated symtab_index_entries. symtab_index_entries is not that
large, it's movable, and it's cheap to move. Making the vector hold
values instead improves cache locality and eliminates many roundtrips
to the heap.
Using the same test as in the previous patch, against the same gdb
inferior, timing improves ~13% further:
~6.0s => ~5.2s (average of 5 runs).
Note that before the .gdb_index C++ifycation patch, we were at ~5.7s.
We're now consistenly better than before.
gdb/ChangeLog
2017-06-12 Pedro Alves <palves@redhat.com>
* dwarf2read.c (mapped_symtab::data): Now a vector of
symtab_index_entry instead of vector of
std::unique_ptr<symtab_index_entry>. All users adjusted to check
whether an element's name is NULL instead of checking whether the
element itself is NULL.
(find_slot): Change return type. Adjust.
(hash_expand, , add_index_entry, uniquify_cu_indices)
(write_hash_table): Adjust.
Using the same test as the previous patch, perf shows GDB spending
over 7% in "free". A substantial number of those calls comes from
insertions in the psyms_seen unordered_set causing lots of rehashing
and recreating buckets. Fix this by computing an estimate of the size
of the set upfront.
Using the same test as in the previous patch, against the same gdb
inferior, timing improves ~8% further:
~6.5s => ~6.0s (average of 5 runs).
gdb/ChangeLog:
2017-06-12 Pedro Alves <palves@redhat.com>
* dwarf2read.c (recursively_count_psymbols): New function.
(write_psymtabs_to_index): Call it to compute number of psyms and
pass estimate size of psyms_seen to unordered_set's ctor.
"perf" shows the unordered_map::emplace call in write_hash_table a bit
high up on profiles. Fix this using the find + insert idiom instead
of going straight to insert.
I tried doing the same to the other unordered_maps::emplace calls in
the file, but saw no performance improvement, so left them be.
With a '-g3 -O2' build of gdb, and:
$ cat save-index.cmd
set $i = 0
while $i < 100
save gdb-index .
set $i = $i + 1
end
$ time ./gdb -data-directory=data-directory -nx --batch -q -x save-index.cmd ./gdb.pristine
I get an improvement of ~7%:
~7.0s => ~6.5s (average of 5 runs).
gdb/ChangeLog:
2017-06-12 Pedro Alves <palves@redhat.com>
* dwarf2read.c (write_hash_table): Check if key already exists
before emplacing.
This avoids having to specify the integer size twice in the same line.
gdb/ChangeLog:
2017-06-12 Pedro Alves <palves@redhat.com>
* dwarf2read.c (data_buf::append_space): Rename to...
(data_buf::grow): ... this, and make private. Adjust all callers.
(data_buf::append_uint): New method.
(add_address_entry, write_one_signatured_type)
(write_psymtabs_to_index): Use it.
There's no real need for all this indirection.
gdb/ChangeLog:
2017-06-12 Pedro Alves <palves@redhat.com>
* dwarf2read.c (file_write(FILE *, const void *, size_t)): Delete.
(file_write (FILE *, const std::vector<Elem>&)): Delete.
(data_buf::file_write): Call ::fwrite directly.
gdb/ChangeLog
2017-06-12 Jan Kratochvil <jan.kratochvil@redhat.com>
Code cleanup: C++ify .gdb_index producer.
* dwarf2read.c: Include <unordered_set> and <unordered_map>.
(MAYBE_SWAP) [WORDS_BIGENDIAN]: Cast to offset_type.
(struct strtab_entry, hash_strtab_entry, eq_strtab_entry)
(create_strtab, add_string): Remove.
(file_write, data_buf): New.
(struct symtab_index_entry): Use std::vector for cu_indices.
(struct mapped_symtab): Use std::vector for data.
(hash_symtab_entry, eq_symtab_entry, delete_symtab_entry)
(create_symbol_hash_table, create_mapped_symtab, cleanup_mapped_symtab):
Remove.
(find_slot): Change return type. Update it to the new data structures.
(hash_expand, add_index_entry): Update it to the new data structures.
(offset_type_compare): Remove.
(uniquify_cu_indices): Update it to the new data structures.
(c_str_view, c_str_view_hasher, vector_hasher): New.
(add_indices_to_cpool): Remove.
(write_hash_table): Update it to the new data structures.
(struct psymtab_cu_index_map, hash_psymtab_cu_index)
(eq_psymtab_cu_index): Remove.
(psym_index_map): New typedef.
(struct addrmap_index_data): Change addr_obstack pointer to data_buf
reference and std::unordered_map for cu_index_htab.
(add_address_entry, add_address_entry_worker, write_address_map)
(write_psymbols): Update it to the new data structures.
(write_obstack): Remove.
(struct signatured_type_index_data): Change types_list to a data_buf
reference and psyms_seen to a std::unordered_set reference.
(write_one_signatured_type, recursively_write_psymbols)
(write_psymtabs_to_index): Update it to the new data structures.
This changes increment_reading_symtab to return a scoped_restore, then
fixes up the users.
gdb/ChangeLog
2017-04-12 Tom Tromey <tom@tromey.com>
* symfile.h (increment_reading_symtab): Update type.
* symfile.c (decrement_reading_symtab): Remove.
(increment_reading_symtab): Return a scoped_restore_tmpl<int>.
* psymtab.c (psymtab_to_symtab): Update.
* dwarf2read.c (dw2_instantiate_symtab): Update.
While the C++ standard says that char16_t and char32_t are unsigned types:
Types char16_t and char32_t denote distinct types with the same size,
signedness, and alignment as uint_least16_t and uint_least32_t,
respectively, in <cstdint>, called the underlying types.
... gdb treats them as signed currently:
(gdb) p (char16_t)-1
$1 = -1 u'\xffff'
There are actually two places in gdb that hardcode these types:
- gdbtypes.c:gdbtypes_post_init, when creating the built-in types,
seemingly used by the "x /s" command (judging from commit 9a22f0d0).
- dwarf2read.c, when reading base types with DW_ATE_UTF encoding
(which is what is used for these types, when compiling for C++11 and
up). Despite the comment, the type created does end up used.
Both places need fixing. But since I couldn't tell why dwarf2read.c
needs to create a new type, I've made it use the per-arch built-in
types instead, so that the types are only created once per arch
instead of once per objfile. That seems to work fine.
While writting the test, I noticed that the C++ language parser isn't
actually aware of these built-in types, so if you try to use them
without a program that uses them, you get:
(gdb) set language c++
(gdb) ptype char16_t
No symbol table is loaded. Use the "file" command.
(gdb) ptype u"hello"
No type named char16_t.
(gdb) p u"hello"
No type named char16_t.
That's fixed by simply adding a couple entries to C++'s built-in types
array in c-lang.c. With that, we get the expected:
(gdb) ptype char16_t
type = char16_t
(gdb) ptype u"hello"
type = char16_t [6]
(gdb) p u"hello"
$1 = u"hello"
gdb/ChangeLog:
2017-04-12 Pedro Alves <palves@redhat.com>
PR c++/21323
* c-lang.c (cplus_primitive_types) <cplus_primitive_type_char16_t,
cplus_primitive_type_char32_t>: New enum values.
(cplus_language_arch_info): Register cplus_primitive_type_char16_t
and cplus_primitive_type_char32_t.
* dwarf2read.c (read_base_type) <DW_ATE_UTF>: If bit size is 16 or
32, use the archtecture's built-in type for char16_t and char32_t,
respectively. Otherwise, fallback to init_integer_type as before,
but make the type unsigned, and issue a complaint.
* gdbtypes.c (gdbtypes_post_init): Make char16_t and char32_t unsigned.
gdb/testsuite/ChangeLog:
2017-04-12 Pedro Alves <palves@redhat.com>
PR c++/21323
* gdb.cp/wide_char_types.c: New file.
* gdb.cp/wide_char_types.exp: New file.
While I was looking at the file, I noticed that this struct could be
nicely converted to a class. As I was progressing, I ended up moving
all state machine actual internal state manipulation to methods of
lnp_state_machine, essentially decoupling DWARF parsing from state
tracking. I also noticed that the lnp_reader_state doesn't really
serve any good use, so that's eliminated in the process.
gdb/ChangeLog:
2017-04-04 Pedro Alves <palves@redhat.com>
* dwarf2read.c (lnp_state_machine): Now a class. Initialize all
data fields, make them private and add "m_" prefixes.
(lnp_state_machine::lnp_state_machine): New ctor.
(record_line, check_line_address, handle_set_discriminator)
(handle_set_address, handle_advance_pc, handle_special_opcode)
(handle_advance_line, handle_set_file, handle_negate_stmt)
(handle_const_add_pc, handle_fixed_advance_pc, handle_copy)
(end_sequence, advance_line): New methods.
(m_gdbarch, m_record_lines_p): New fields.
(lnp_reader_state): Delete.
(dwarf_record_line): Rename to ...
(lnp_state_machine::record_line): ... adjust.
(init_lnp_state_machine): Delete.
(lnp_state_machine::lnp_state_machine): New.
(check_line_address): Rename to ...
(lnp_state_machine::check_line_address): This.
(dwarf_decode_lines_1): Remove reference to "reader_state".
Adjust lnp_state_machine having a non-default ctor. Use bool.
State machine internal state manipulation moved to
lnp_state_machine methods.
A while ago, back when GDB was a C program, the sect_offset and
cu_offset types were made structs in order to prevent incorrect mixing
of those offsets. Now that we require C++11, we can make them
integers again, while keeping the safety, by exploiting "enum class".
We can add a bit more safety, even, by defining operators that the
types _should_ support, helping making the suspicious uses stand out
more.
Getting at the underlying type is done with the new to_underlying
function added by the previous patch, which also helps better spot
where do we need to step out of the safety net. Mostly, that's around
parsing the DWARF, and when we print the offset for complaint/debug
purposes. But there are other occasional uses.
Since we have to define the sect_offset/cu_offset types in a header
anyway, I went ahead and generalized/library-fied the idea of "offset"
types, making it trivial to add more such types if we find a use. See
common/offset-type.h and the DEFINE_OFFSET_TYPE macro.
I needed a couple generaly-useful preprocessor bits (e.g., yet another
CONCAT implementation), so I started a new common/preprocessor.h file.
I included units tests covering the "offset" types API. These are
mostly compile-time tests, using SFINAE to check that expressions that
shouldn't compile (e.g., comparing unrelated offset types) really are
invalid and would fail to compile. This same idea appeared in my
pending enum-flags revamp from a few months ago (though this version
is a bit further modernized compared to what I had posted), and I plan
on reusing the "check valid expression" bits added here in that
series, so I went ahead and defined the CHECK_VALID_EXPR macro in its
own header -- common/valid-expr.h. I think that's nicer regardless.
I was borderline between calling the new types "offset" types, or
"index" types, BTW. I stuck with "offset" simply because that's what
we're already calling them, mostly.
gdb/ChangeLog:
2017-04-04 Pedro Alves <palves@redhat.com>
* Makefile.in (SUBDIR_UNITTESTS_SRCS): Add
unittests/offset-type-selftests.c.
(SUBDIR_UNITTESTS_OBS): Add offset-type-selftests.o.
* common/offset-type.h: New file.
* common/preprocessor.h: New file.
* common/traits.h: New file.
* common/valid-expr.h: New file.
* dwarf2expr.c: Include "common/underlying.h". Adjust to use
sect_offset and cu_offset strong typedefs throughout.
* dwarf2expr.h: Adjust to use sect_offset and cu_offset strong
typedefs throughout.
* dwarf2loc.c: Include "common/underlying.h". Adjust to use
sect_offset and cu_offset strong typedefs throughout.
* dwarf2read.c: Adjust to use sect_offset and cu_offset strong
typedefs throughout.
* gdbtypes.h: Include "common/offset-type.h".
(cu_offset): Now an offset type (strong typedef) instead of a
struct.
(sect_offset): Likewise.
(union call_site_parameter_u): Rename "param_offset" field to
"param_cu_off".
* unittests/offset-type-selftests.c: New file.
This should help catch mistakes related to mixing the 1-based DWARF
indexes with 0-based std::vector indexes, since the new types do not
implicitly convert to anything.
The change in read_formatted_entries relates to the fact that doing
the seemingly simpler:
- uintp = &fe.dir_index;
+ uintp = (unsigned int *) &fe.dir_index;
would be undefined C/C++. So to address that, I made the function
extract the form before assigning to the file_entry. It felt natural
to use gdb::optional for "do I have this value", and this is what
motivated the previous patch that added the missing observer methods
to gdb::optional.
gdb/ChangeLog:
2017-04-04 Pedro Alves <palves@redhat.com>
* common/underlying.h: New file.
* dwarf2read.c: Include "common/gdb_optional.h" and
"common/underlying.h".
(dir_index, file_name_index): New types.
(file_entry): Use them.
(file_entry::include): Use to_underlying.
(line_header::add_file_name): Use dir_index.
(read_formatted_entries): Use gdb::optional. Read form before
writting to file_entry.
(dwarf_decode_line_header): Use dir_index.
(lnp_state_machine::current_file): Use to_underlying.
(lnp_state_machine::file): Change type to file_name_index.
(dwarf_record_line): Use to_underlying.
(init_lnp_state_machine): Use file_name_index.
(dwarf_decode_lines_1): Use dir_index and file_name_index.
This starts off as replacing a couple custom open coded vector
implementations in the file with std::vector, and then the rest falls
off of that. I.e., use new/delete instead of XCNEW/xfree, add
ctors/dtors/initializers where appropriate. And then use
std::unique_ptr instead of cleanups. Some functions became methods,
and in a couple spots, some single-use callback functions that would
have to be tweaked anyway are converted to lambdas instead.
gdb/ChangeLog:
2017-04-04 Pedro Alves <palves@redhat.com>
* dwarf2read.c (struct file_entry): Add ctors, and initialize all
fields.
(line_header): Initialize all data fields. Change type of
standard_opcode_lengths to std::unique_ptr<unsigned char[]>.
Change type of include_dirs to std::vector<const char *>. Remove
num_include_dirs, include_dirs_size. Change type of file_names to
std::vector<file_entry>. Remove num_file_names, file_names_size.
(line_header::line_header): New.
(line_header::add_include_dir, line_header::add_file_name): New
methods.
(line_header::include_dir_at): Remove NULL check.
(line_header::file_name_at): Add const overload.
(line_header_up): New unique_ptr typedef.
(dw2_get_file_names_reader): Use line_header_up. Adjust to use
std::vector. Remove free_line_header call.
(dwarf2_build_include_psymtabs): Use line_header_up. Remove
free_line_header call.
(free_cu_line_header): Delete.
(handle_DW_AT_stmt_list, handle_DW_AT_stmt_list)
(setup_type_unit_groups): Use line_header_up instead of cleanups.
Adjust to use std::vector.
(free_line_header): Delete.
(free_line_header_voidp): Use delete.
(add_include_dir): Replace with ...
(line_header::add_include_dir): ... this method. Use std::vector.
(add_file_name): Replace with ...
(line_header::add_file_name): ... this method. Use std::vector.
(add_include_dir_stub): Delete.
(read_formatted_entries): Remove memset.
(dwarf_decode_line_header): Return a line_header_up instead of a
raw pointer. Remove cleanup handling. Pass lambdas to
read_formatted_entries. Adjust to use line_header methods.
(dwarf_decode_lines_1): Adjust to use line_header methods.
(dwarf_decode_lines, file_file_name, file_full_name): Adjust to
use std::vector.
Multiple places in dwarf2read.c open code 1-based to 0-based index
conversion and check for out of bounds accesses to lh->include_dirs
and lh->file_names. This commit factors those out to a couple methods
and uses them throughout.
gdb/ChangeLog:
2017-03-27 Pedro Alves <palves@redhat.com>
* dwarf2read.c (file_entry) <dir_index>: Add comment.
(file_entry::include_dir): New method.
(line_header::include_dir_at, line_header::file_name_at): New
methods.
(setup_type_unit_groups, setup_type_unit_groups)
(psymtab_include_file_name): Simplify using the new methods.
(lnp_state_machine) <the_line_header>: New field.
<file>: Add comment.
(lnp_state_machine::current_file): New method.
(dwarf_record_line): Simplify using the new methods.
(init_lnp_state_machine): Initialize the "the_line_header" field.
(dwarf_decode_lines_1, dwarf_decode_lines, file_file_name):
Simplify using the new methods.
Eliminates several uses of cleanups.
Tested on x86_64 Fedora 23 with Python 2 and 3.
gdb/ChangeLog
2017-03-27 Pedro Alves <palves@redhat.com>
* dwarf2read.c (struct file_and_directory): New.
(dwarf2_get_dwz_file): Adjust to use std::string.
(dw2_get_file_names_reader): Adjust to use file_and_directory.
(find_file_and_directory): Adjust to return a file_and_directory
object.
(read_file_scope): Adjust to use file_and_directory. Remove
make_cleanup/do_cleanups calls.
(open_and_init_dwp_file): Adjust to use std::string. Remove
make_cleanup/do_cleanups calls.
* python/python.c (do_start_initialization): Adjust to ldirname
returning a std::string.
* utils.c (ldirname): Now returns a std::string.
* utils.h (ldirname): Change return type to std::string.
* xml-syscall.c (xml_init_syscalls_info): Adjust to ldirname
returning a std::string.
* xml-tdesc.c (file_read_description_xml): Likewise.
gdb was segfaulting during backtrace on a binary here, where
fe->dir_index parsed from the DWARF info was seen to access beyond the
provided include_dirs array.
This commit bounds the access to entries actually written to the
array, and was verified to output the backtrace correctly.
gdb/ChangeLog:
* dwarf2read.c (setup_type_unit_groups): Ensure dir_index doesn't
reference beyond the 'lh->include_dirs' array before accessing to
it.
(psymtab_include_file_name): Likewise.
(dwarf_decode_lines_1): Likewise.
(dwarf_decode_lines): Likewise.
(file_file_name): Likewise.
Make gdb DWARF reader understand the DW_TAG_rvalue_reference type tag. Handling
of this tag is done in the existing read_tag_reference_type() function, to
which we add a new parameter representing the kind of reference type
(lvalue vs rvalue).
gdb/ChangeLog
PR gdb/14441
* dwarf2read.c (process_die, read_type_die_1): Handle the
DW_TAG_rvalue_reference_type DIE.
(read_tag_reference_type): Add new parameter "refcode".
Parameterize lookup_reference_type() and make_reference_type() by the kind of
reference type we want to look up. Create two wrapper functions
lookup_{lvalue,rvalue}_reference_type() for lookup_reference_type() to simplify
the API. Change all callers to use the new API.
gdb/Changelog
PR gdb/14441
* dwarf2read.c (read_tag_reference_type): Use
lookup_lvalue_reference_type() instead of lookup_reference_type().
* eval.c (evaluate_subexp_standard): Likewise.
* f-exp.y: Likewise.
* gdbtypes.c (make_reference_type, lookup_reference_type):
Generalize with rvalue reference types.
(lookup_lvalue_reference_type, lookup_rvalue_reference_type): New
convenience wrappers for lookup_reference_type().
* gdbtypes.h (make_reference_type, lookup_reference_type): Add a
reference kind parameter.
(lookup_lvalue_reference_type, lookup_rvalue_reference_type): Add
wrappers for lookup_reference_type().
* guile/scm-type.c (gdbscm_type_reference): Use
lookup_lvalue_reference_type() instead of lookup_reference_type().
* guile/scm-value.c (gdbscm_value_dynamic_type): Likewise.
* parse.c (follow_types): Likewise.
* python/py-type.c (typy_reference, typy_lookup_type): Likewise.
* python/py-value.c (valpy_get_dynamic_type, valpy_getitem):
Likewise.
* python/py-xmethods.c (gdbpy_get_xmethod_result_type)
(gdbpy_invoke_xmethod): Likewise.
* stabsread.c: Provide extra argument to make_reference_type()
call.
* valops.c (value_ref, value_rtti_indirect_type): Use
lookup_lvalue_reference_type() instead of lookup_reference_type().
Recently I fixed a bug that caused a DW_OP_implicit_pointer with non-zero
offset into a DW_OP_implicit_value to be handled incorrectly on big-endian
targets. GDB ignored the offset and copied the wrong bytes:
https://sourceware.org/ml/gdb-patches/2017-01/msg00251.html
But there is still a similar issue when a DW_OP_implicit_pointer points
into a DW_OP_stack_value instead; and again, the offset is ignored. There
is an important difference, though: While implicit values are treated like
blocks of data and anchored at the lowest-addressed byte, stack values
traditionally contain integer numbers and are anchored at the *least
significant* byte. Also, stack values do not come in varying sizes, but
are cut down appropriately when used. Thus, on big-endian targets the
scenario looks like this (higher addresses shown right):
|<- - - - - Stack value - - - - - - ->|
| |
|<- original object ->|
|
| offset ->|####|
^^^^
de-referenced
implicit pointer
(Note how the original object's size influences the position of the
de-referenced implicit pointer within the stack value. This is not the
case for little-endian targets, where the original object starts at offset
zero within the stack value.)
This patch implements the logic indicated in the above diagram and adds an
appropriate test case. A new function dwarf2_fetch_die_type_sect_off is
added; it is used for retrieving the original object's type, so its size
can be determined. That type is passed to dwarf2_evaluate_loc_desc_full
via a new parameter.
gdb/ChangeLog:
* dwarf2loc.c (indirect_synthetic_pointer): Get data type of
pointed-to DIE and pass it to dwarf2_evaluate_loc_desc_full.
(dwarf2_evaluate_loc_desc_full): New parameter subobj_type; rename
byte_offset to subobj_byte_offset. Fix the handling of
DWARF_VALUE_STACK on big-endian targets when coming via an
implicit pointer.
(dwarf2_evaluate_loc_desc): Adjust call to
dwarf2_evaluate_loc_desc_full.
* dwarf2loc.h (dwarf2_fetch_die_type_sect_off): New declaration.
* dwarf2read.c (dwarf2_fetch_die_type_sect_off): New function.
gdb/testsuite/ChangeLog:
* lib/dwarf.exp: Add support for DW_OP_implicit_pointer.
* gdb.dwarf2/nonvar-access.exp: Add test for stack value location
and implicit pointer into such a location.
gcc-6.3.1-1.fc25.x86_64
dwarf2read.c: In function ‘void create_debug_type_hash_table(dwo_file*, dwarf2_section_info*, htab*&, rcuh_kind)’:
dwarf2read.c:4776:32: error: ‘header.comp_unit_head::type_offset_in_tu.cu_offset::cu_off’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
dwarf2read.c:4816:21: error: ‘header.comp_unit_head::signature’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
cc1plus: all warnings being treated as errors
gdb/ChangeLog
2017-02-26 Jan Kratochvil <jan.kratochvil@redhat.com>
* dwarf2read.c (create_debug_type_hash_table): Initialize
header.signature and header.type_offset_in_tu.
I wanted to pass a lambda to iterate_over_symtabs (see following
patch), so I converted it to function_view, and then the rest is
cascaded from that.
This gets rid of a bunch of single-use callback functions and
corresponding manually managed callback capture types
(add_partial_datum, search_symbols_data, etc.) in favor of letting the
compiler generate them for us by using lambdas with a capture. In a
couple cases, it was more natural to convert the existing function
callbacks to function objects (i.e., operator(), e.g.,
decode_compound_collector).
gdb/ChangeLog:
2017-02-23 Pedro Alves <palves@redhat.com>
* ada-lang.c: Include "common/function-view.h".
(ada_iterate_over_symbols): Adjust to use function_view as
callback type.
(struct add_partial_datum, ada_complete_symbol_matcher): Delete.
(ada_make_symbol_completion_list): Use a lambda.
(ada_exc_search_name_matches): Delete.
(name_matches_regex): New.
(ada_add_global_exceptions): Use a lambda and name_matches_regex.
* compile/compile-c-support.c: Include "common/function-view.h".
(print_one_macro): Change prototype to accept a ui_file pointer.
(write_macro_definitions): Use a lambda.
* dwarf2read.c: Include "common/function-view.h".
(dw2_map_expand_apply, dw2_map_symtabs_matching_filename)
(dw2_expand_symtabs_matching): Adjust to use function_view as
callback type.
* language.h: Include "common/function-view.h".
(struct language_defn) <la_iterate_over_symbols>: Adjust to use
function_view as callback type.
(LA_ITERATE_OVER_SYMBOLS): Remove DATA parameter.
* linespec.c: Include "common/function-view.h".
(collect_info::add_symbol): New method.
(struct symbol_and_data_callback, iterate_inline_only, struct
symbol_matcher_data, iterate_name_matcher): Delete.
(iterate_over_all_matching_symtabs): Adjust to use function_view
as callback type and lambdas.
(iterate_over_file_blocks): Adjust to use function_view as
callback type.
(decode_compound_collector): Now a class with private fields.
(decode_compound_collector::release_symbols): New method.
(collect_one_symbol): Rename to...
(decode_compound_collector::operator()): ... this and adjust.
(lookup_prefix_sym): decode_compound_collector construction bits
move to decode_compound_collector ctor. Pass the
decode_compound_collector object directly as callback. Remove
cleanups and use decode_compound_collector::release_symbols
instead.
(symtab_collector): Now a class with private fields.
(symtab_collector::release_symtabs): New method.
(add_symtabs_to_list): Rename to...
(symtab_collector::operator()): ... this and adjust.
(collect_symtabs_from_filename): symtab_collector construction
bits move to symtab_collector ctor. Pass the symtab_collector
object directly as callback. Remove cleanups and use
symtab_collector::release_symtabs instead.
(collect_symbols): Delete.
(add_matching_symbols_to_info): Use lambdas.
* macrocmd.c (print_macro_callback): Delete.
(info_macro_command): Use a lambda.
(info_macros_command): Pass print_macro_definition as callable
directly.
(print_one_macro): Remove 'ignore' parameter.
(macro_list_command): Adjust.
* macrotab.c (macro_for_each_data::fn): Now a function_view.
(macro_for_each_data::user_data): Delete field.
(foreach_macro): Adjust to call the function_view.
(macro_for_each): Adjust to use function_view as callback type.
(foreach_macro_in_scope): Adjust to call the function_view.
(macro_for_each_in_scope): Adjust to use function_view as callback
type.
* macrotab.h: Include "common/function-view.h".
(macro_callback_fn): Declare a prototype instead of a pointer.
Remove "user_data" parameter.
(macro_for_each, macro_for_each_in_scope): Adjust to use
function_view as callback type.
* psymtab.c (partial_map_expand_apply)
(psym_map_symtabs_matching_filename, recursively_search_psymtabs):
Adjust to use function_view as callback type and to return bool.
(psym_expand_symtabs_matching): Adjust to use function_view as
callback types.
* symfile-debug.c (debug_qf_map_symtabs_matching_filename): Adjust
to use function_view as callback type and to return bool.
(debug_qf_expand_symtabs_matching): Adjust to use function_view as
callback types.
* symfile.c (expand_symtabs_matching): Adjust to use function_view
as callback types.
* symfile.h: Include "common/function-view.h".
(expand_symtabs_file_matcher_ftype)
(expand_symtabs_symbol_matcher_ftype)
(expand_symtabs_exp_notify_ftype): Remove "data" parameter and
return bool.
(quick_symbol_functions::map_symtabs_matching_filename)
(quick_symbol_functions::expand_symtabs_matching): Adjust to use
function_view as callback type and return bool.
(expand_symtabs_matching): Adjust to use function_view as callback
type.
(maintenance_expand_name_matcher)
(maintenance_expand_file_matcher): Delete.
(maintenance_expand_symtabs): Use lambdas.
* symtab.c (iterate_over_some_symtabs): Adjust to use
function_view as callback types and return bool.
(iterate_over_symtabs): Likewise. Use unique_xmalloc_ptr instead
of a cleanup.
(lookup_symtab_callback): Delete.
(lookup_symtab): Use a lambda.
(iterate_over_symbols): Adjust to use function_view as callback
type.
(struct search_symbols_data, search_symbols_file_matches)
(search_symbols_name_matches): Delete.
(search_symbols): Use a pair of lambdas.
(struct add_name_data, add_macro_name, symbol_completion_matcher)
(symtab_expansion_callback): Delete.
(default_make_symbol_completion_list_break_on_1): Use lambdas.
* symtab.h: Include "common/function-view.h".
(iterate_over_some_symtabs): Adjust to use function_view as
callback type and return bool.
(iterate_over_symtabs): Adjust to use function_view as callback
type.
(symbol_found_callback_ftype): Remove 'data' parameter and return
bool.
(iterate_over_symbols): Adjust to use function_view as callback
type.
On some Fedora 23 systems an internal error has been printed.
gdb/ChangeLog
2017-02-21 Jan Kratochvil <jan.kratochvil@redhat.com>
* dwarf2read.c (dwarf2_record_block_ranges): Add forgotten BASEADDR.
gcc-4.8.5-11.el7.x86_64
dwarf2read.c: In function ‘pc_bounds_kind dwarf2_get_pc_bounds(die_info*, CORE_ADDR*, CORE_ADDR*, dwarf2_cu*, partial_symtab*)’:
dwarf2read.c:12134:7: error: ‘range_end’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
dwarf2read.c:12133:7: error: ‘range_beginning’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
gdb/ChangeLog
2017-02-21 Jan Kratochvil <jan.kratochvil@redhat.com>
* dwarf2_rnglists_process: Initialize range_beginning and range_end.
DWARF-5 has new form DW_FORM_data16. The problem is that GDB cannot pass
16-byte constant as a constant value as that would require GDB to use GCC
extension __int128.
Formerly such data was coded as DW_FORM_block* so GDB still decodes
DW_FORM_data16 like DW_FORM_block*.
gdb/ChangeLog
2017-02-20 Jan Kratochvil <jan.kratochvil@redhat.com>
* dwarf2read.c (skip_one_die, read_attribute_value)
(dwarf2_const_value_attr, dump_die_shallow)
(dwarf2_get_attr_constant_value, dwarf2_fetch_constant_bytes)
(skip_form_bytes, attr_form_is_constant): Handle DW_FORM_data16.
gdb/testsuite/ChangeLog
2017-02-20 Jan Kratochvil <jan.kratochvil@redhat.com>
* gdb.dwarf2/formdata16.c: New file.
* gdb.dwarf2/formdata16.exp: New file.
* lib/dwarf.exp (Dwarf): Add DW_FORM_data16.