The stabs reading code in google-breakpad incorrectly assumes that the
stabs data is a single compilation unit. Specifically, it ignores
N_UNDF stabs and assumes that all string indices are relative to the
beginning of the .stabstr section.
This is true when linking with the GNU linker by default, because the
GNU linker optimizes stabs debug info. The gold linker does not do
this optimization. It can be disabled when using the GNU linker with
the --traditional-format command line option.
For more details of the problem, see:
http://sourceware.org/bugzilla/show_bug.cgi?id=10338http://code.google.com/p/google-breakpad/issues/detail?id=359
This patch adds unit tests that reproduce the failure, and fixes the
stabs parser.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@490 4c0a9323-5329-0410-9bdc-e9ce6186880e
Add a new member function to dwarf2reader::Dwarf2Handler,
ProcessAttributeReference, for reporting attribute values that are
references to other DIEs. This handler member function always receives
an absolute offset (that is, relative to the start of the .debug_info
section, not to the start of the compilation unit), regardless of the
form the attribute uses. (Some forms are CU-relative, some are
absolute.)
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@482 4c0a9323-5329-0410-9bdc-e9ce6186880e
Previous patches added unit tests for the STABS parser and the
Breakpad symbol file writer; this adds unit tests for the "dumper"
class that sits between them, receiving data from the parser and
handing it to the writer. So now the whole pathway has coverage.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@467 4c0a9323-5329-0410-9bdc-e9ce6186880e
Adjust Module's interface a bit to facilitate testing:
- Make AssignSourceIds something a client can call --- it's perfectly
well-defined, so this is an okay change.
- Add GetFunctions, GetFiles and FindExistingfile member functions,
which the test harness will use to get results to examine.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@466 4c0a9323-5329-0410-9bdc-e9ce6186880e
A FUNC record's parameter size is also hexadecimal, and all values are
64 bits wide.
A line record's address and size are 64 bits wide.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@465 4c0a9323-5329-0410-9bdc-e9ce6186880e
Fix typos.
For CompilationUnit::Start, I was confused by the '-' in the original
comment, taking it for a parenthetic clause marker, assuming an
implicit "of the next compilation unit" at the end of the sentence.
The comments should refer to the ".debug_info" section, not the
"debug_info" section. The latter is not the section name actually used
on any system (ELF or Mach-O), and the former is the name prescribed
by the DWARF spec.
Some of the comments for ProcessAttribute* member functions claim that
OFFSET is from the start of the compilation unit, but that's not so:
the code has always passed an offset relative to the start of the
.debug_info section.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@453 4c0a9323-5329-0410-9bdc-e9ce6186880e
src/common/dwarf/dwarf2reader.cc uses the old-fashioned <stdio.h>
facilities to report errors. Ideally, we would add a 'Warning' message
to the handler and make the client responsible for dealing with the
errors, but this at least allows us to compile.
Ubuntu 9.10 uses GCC 4.4.1; under older versions of GCC, this wasn't a
problem, probably because stdio.h was being brought in inadvertently
somewhere else.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@449 4c0a9323-5329-0410-9bdc-e9ce6186880e
It seems that a use of the <stdint.h> type uintptr_t has crept into
the DWARF parser. This defines a workaround for the GNU compilers
(tested on both Mac and Linux) which will raise an error if it doesn't
work.
My personal preference would be just to assume that the <stdint.h>
header is available and use the standard types everywhere, but 1) that
would be a large change, likely to make merges with the other branches
of the DWARF parser more difficult, and 2) it would make it quite
difficult to build under Microsoft Visual Studio, which doesn't have
the <stdint.h> header; Microsoft has said they have no plans to
provide it, as they would rather "focus their efforts" on C++ and
.NET.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@448 4c0a9323-5329-0410-9bdc-e9ce6186880e
Building on Ubuntu 9.10 with the distributed compiler (GCC 4.4.1), we get
warnings like the following:
guid_creator.cc:56: warning: dereferencing type-punned pointer will break strict-aliasing rules
It doesn't matter in this case, but there's no crying need to use
reinterpret casts in an endian-dependent way when there are plenty of
well-defined ways to get the same effect.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@447 4c0a9323-5329-0410-9bdc-e9ce6186880e
Move the DWARF parser, and the functioninfo.cc DWARF consumer, from
src/common/mac/dwarf to src/commmon/dwarf, so that it can be shared
between the Mac and Linux dumpers.
Fix up #include directives, multiple inclusion protection macros, and
Xcode build files.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@446 4c0a9323-5329-0410-9bdc-e9ce6186880e
The test system is based on Google C++ Testing Framework and the
Google C++ Mocking Framework.
This includes a parser that turns human-readable input files ("mock
stabs") into .stab and .stabstr section contents, which we can then
pass to a StabsReader instance, using a handler object written with
GoogleMock. The 'make check' target in src/tools/linux/dump_syms runs
this.
The supplied input file is pretty small, but I've done coverage
testing, and it does cover the parser.
I thought the mock stabs parser would be less elaborate than it turned
out to be. Lesson learned.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@444 4c0a9323-5329-0410-9bdc-e9ce6186880e
If the input passed to a StabsReader instance contains a compilation
unit whose first entry is an N_SO with no name, the parser enters an
infinite loop. Since such entries mark the end of a compilation unit,
ProcessCompilationUnit should skip them.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@443 4c0a9323-5329-0410-9bdc-e9ce6186880e
The StabsHandler class should not provide a fallback definition for
its Warning member function that just throws away warning messages.
It should require the consumer to provide an appropriate definition.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@442 4c0a9323-5329-0410-9bdc-e9ce6186880e
Modern GNU compilers warn about the #inclusion of <ext/hash_map>; that
container is deprecated, and code should use <tr1/unordered_map>
instead. However, to stay within the boundaries of C++ '98, it's
probably fine just to use plain old std::map.
Breakpad uses hash_map in three cases:
o The DWARF reader's SectionMap type maps object file section names to
data. This map is consulted once per section kind per DWARF
compilation unit; it is not performance-critical.
o The Mac dump_syms tool uses it to map machine architectures to
section maps in Universal binaries. It's hard to imagine there
ever being more than two entries in such a map.
o The processor's BasicSourceLineResolver uses a hash_map to map file
numbers to file names. This is the map that will probably have the
most entries, but it's only accessed once per frame, after we've
found the frame's line entry.
a=jimblandy
r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@393 4c0a9323-5329-0410-9bdc-e9ce6186880e
Fix some typos and references to member functions that didn't make the
final cut.
a=jimblandy
r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@381 4c0a9323-5329-0410-9bdc-e9ce6186880e
src/linux/common/module.h defines a new class, google_breakpad::Module,
that can represent the contents of a breakpad symbol file. Module::Write
writes a well-formed symbol file to the given stream.
src/linux/common/dump_symbols.cc can now lose its symbol-file-writing
code, and change DumpStabsHandler to populate a Module object, rather
than the old SymbolInfo/SourceFileInfo/... collection of types.
The code to compute function and line sizes, even in the absence of
reliable size data in STABS, is moved into a new Finalize method of
DumpStabsHandler, which is responsible for completing the Module's
contents.
a=jimblandy
r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@380 4c0a9323-5329-0410-9bdc-e9ce6186880e
With this patch, dump_symbols.cc no longer knows about the details of
the STABS debugging format; that is handled by the StabsReader class.
dump_symbols.cc provides a subclass of StabsHandler that builds
dump_symbols' own representation of the data.
a=jimblandy
r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@378 4c0a9323-5329-0410-9bdc-e9ce6186880e
Because the actual N_FUN strings in the .stabstr section contain type
information after the mangled name, representing this information
using a pointer into .stabstr, while efficient with memory, makes the
FuncInfo data structure STABS-specific: one must know the details of a
STABS N_FUN string's syntax to interpret FuncInfo::name. This patch
removes this STABS dependency from the data structure, and moves us
closer to having an appropriate structure for representing unified
STABS and DWARF data.
a=jimblandy
r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@375 4c0a9323-5329-0410-9bdc-e9ce6186880e
In STABS, if one function's line number information contains an N_SOL
entry to switch to a new source file, then the next function's line
data should pick up in the same source file where the prior function
left off. However, the Linux dumper restarts each function in the
compilation unit's main source file. This patch fixes that, so that
the output attributes the lines in subsequent functions to the correct
source files.
a=jimblandy
r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@373 4c0a9323-5329-0410-9bdc-e9ce6186880e
Let LineInfo structures point directly to their SourceLineInfo
structures, rather than holding the index of the file's name in the
.stabstr section in the early phases, and then later the holding
source_id of the file.
This is another step in the process of moving STABS-specific values
out of the types that represent the breakpad symbol data. When we're
done, the non-STABS structures will be something that we can populate
with both STABS and DWARF data --- or at least it will be more easily
replaced with such.
a=jimblandy
r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@371 4c0a9323-5329-0410-9bdc-e9ce6186880e
STABS information introduces a compilation unit with an N_SO entry
whose address is the start address of the file and whose string is the
name of the compilation unit's main source file. However, STABS
entries can only hold one address, so STABS indicates the compilation
unit's ending address with an N_SO entry whose name is empty.
Currently, the dumper's data structures simply create SourceFileInfo
structures with empty names for these end-of-unit N_SO entries. We
want to remove STABS-specific characteristics from these structures so
that we can replace them with an input-format-independent structure.
This moves end-of-compilation-unit addresses out of the symbol table
structure, and into their own list of boundary addresses.
a=jimblandy
r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@369 4c0a9323-5329-0410-9bdc-e9ce6186880e
Use a list of pointers to SourceFileInfo structures, not a list of the
structures themselves. This is preparation for a subsequent patch
which makes the data structures less STABS-specific.
This patch introduces a memory leak. If an included file is
referenced only by line entries for functions that LoadFuncSymbols
elected to omit from the func_info list, then its SourceFileInfo
structure is leaked when we destroy the name_to_file map. This leak
is fixed in a subsequent patch by letting the map of files by name own
the file objects.
a=jimblandy
r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@368 4c0a9323-5329-0410-9bdc-e9ce6186880e
Replace the sorted lists of files and functions with an array of
boundary addresses. This replaces CompareAddress with the default
comparison, and SortByAddress and NextAddress with the stock STL sort
and upper_bound algorithms, losing ~50 lines of code.
a=jimblandy
r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@367 4c0a9323-5329-0410-9bdc-e9ce6186880e
In NextAddress, check both the file list and the function list for the
nearest boundary. Don't assume that, if we find any bounding entry in
the function list, that must be the nearest thing.
A=jimblandy
R=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@365 4c0a9323-5329-0410-9bdc-e9ce6186880e
The current arrangement would produce needless warnings if
WriteSymbolFile were ever used twice in the same program invocation.
Even if it weren't wrong, it's unnecessary, and local non-const static
variables require extra care when reading to be sure of their effect.
A=jimblandy
R=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@363 4c0a9323-5329-0410-9bdc-e9ce6186880e
With this patch, the time required to generate Breakpad symbols for
Firefox's libxul.so on a MacBook Pro 3,1 drops from 32s to 2s.
I verified that this patch had no effect on the output of dump_syms
when applied to firefox-bin and its libraries when built with -gstabs+.
A=jimblandy
R=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@362 4c0a9323-5329-0410-9bdc-e9ce6186880e
