Other DWARF sections still need to be done, as well as 64-bit offsets in DataExtractor also needed (as described DebugFrame).
Details
Diff Detail
- Repository
- rL LLVM
Event Timeline
Could you include some test cases?
Since LLVM cannot produce DWARF64 yet (so far as I know) you may need to
generate them with GCC (you'll find other binary tests in
test/DebugInfo/Inputs/dwarfdump-* (with corresponding tests like
test/DebugInfo/dwarfdump-test.test)). Please include comments about which
version of GCC and what flags were used to generate the binaries.
I compiled the existing dwarfdump-inl-test.cc for FreeBSD mips64:
% g++ --version
g++ (GCC) 4.2.1 20070831 patched [FreeBSD]
Copyright (C) 2007 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
% g++ -g dwarfdump-inl-test.cc -o dwarfdump-inl-test.freebsd-elf-mips64
current state:
% bin/llvm-dwarfdump --debug-dump=line dwarfdump-inl-test.freebsd-elf-mips64 dwarfdump-inl-test.freebsd-elf-mips64: file format ELF64-unknown .debug_line contents: %
with this patch:
% bin/llvm-dwarfdump --debug-dump=line dwarfdump-inl-test.freebsd-elf-mips64 dwarfdump-inl-test.freebsd-elf-mips64: file format ELF64-unknown .debug_line contents: Line table prologue: total_length: 0x00000212 version: 2 prologue_length: 0x000001ab min_inst_length: 1 default_is_stmt: 1 line_base: -5 line_range: 14 opcode_base: 13 ...
this only outputs the first CompileUnit though, perhaps because the parsing fails in another section in DWARFContext::parseCompileUnits(). Is it reasonable to add this dwarfdump-inl-test.freebsd-elf-mips64 binary and a basic test, fleshing it out when DWARF64 parsing for other sections is added?
(the test file is 15638 bytes -- larger than the clang-compiled version, due to the lack of -gline-tables-only)
Is this actually using 64-bit dwarf or is it dwarf with 64-bit
pointers? If it's the former ... then how big is the testcase? It
really shouldn't be creating dwarf64 unless the references in the
object are large (not the size of a pointer) unless this is dwarf2.
-eric
This patch addresses the output of GCC and gas on FreeBSD mips64
# gcc --version gcc (GCC) 4.2.1 20070831 patched [FreeBSD] # as --version GNU assembler 2.17.50 [FreeBSD] 2007-07-03
This toolchain generally outputs dwarf2 but uses the 64-bit format from dwarf3. This is the format that uses 0xffffffff in the initial length field, followed by a 64-bit length, and 64-bit offsets throughout. Equivalent support already exists in DWARFDebugFrame.cpp.
My testcase binary is tiny; 64-bit is not necessary. The (older) version of GCC and gas in FreeBSD output this format for some reason; I do not know if more recent versions still do.
for whatever reason the toolchain insists on
That's really a shame. For dwarf3 it shouldn't generate 64-bit dwarf
unless the offsets in the file require 64-bits to represent. For
dwarf2 it's based on pointer size. Seems like a bug. Which version of
gcc is this?
-eric
The version was in the Phabricator review or email :-)
gcc 4.2.1, gas "2.17.50"
I agree that this is a shame, but it's what we've got. I haven't investigated in detail, but quick Googling suggests that the GNU toolchain switched n64 Linux to DWARF32 in 2006; I did not find evidence that the same change was made for FreeBSD.
It appears the GNU toolchain was changed to default to dwarf32 on all MIPS platforms sometime later, and I will make a similar change to the FreeBSD in-tree toolchain. That said, we still have MIPS binaries and libraries built with dwarf64 data today.
FreeBSD in-tree toolchain updated in svn r256859:
http://svnweb.freebsd.org/base?view=revision&revision=256859
Shall I still proceed with this patch? We (a) have objects with 64-bit dwarf .debug_line already due to the silly default in older toolchains, and (b) may eventually encounter an object that actually requires this support.
The patch itself is ok, though it needs a few things:
a) comments on what is and is not handled for dwarf64. The compile unit header is probably the best place to document that there's at least some partial support for reading line tables.
b) testcase. I'm ok with adding the testcase and a comment on it that dwarf64 parsing isn't very good at this point and that the only thing that's expected to work are the line tables. You can, for now, just dump the line table in the testcase.
Added comments and a basic test
(Comments are in DWARFContext.h as that seemed to be a more clear reference to the individual classes.)
One comment, otherwise it looks ok.
lib/DebugInfo/DWARFDebugLine.cpp | ||
---|---|---|
220 ↗ | (On Diff #5096) | Not sure I understand the error here. |
lib/DebugInfo/DWARFDebugLine.cpp | ||
---|---|---|
220 ↗ | (On Diff #5096) | gas has a bug that produces an incorrect prologue_length for 64-bit DWARF. As it turns out we can handle that condition, so we can still return true. I fixed the underlying bug in FreeBSD here: Maybe just drop the message? |
It just fell off my radar during the previous iterations. I subsequently committed an improved version of the change to LLDB - that's in http://llvm-reviews.chandlerc.com/D2007 .
We'll want the equivalent change here too, but I think it makes sense to bring in this by itself since it's already done. It wasn't a trivial rebase against head so I'll commit if I can get a review of the new diff.
test/DebugInfo/dwarfdump-64-bit-dwarf.test | ||
---|---|---|
1 ↗ | (On Diff #10244) | I named it this only because it was built from dwarfdump-inl-test.cc The actual source isn't really relevant though, so I can just call it dwarfdump.elf-mips64-64-bit-dwarf if that's preferable. |
Sorry this slipped through the cracks for so long. I'll commit shortly, but have uploaded a rebased version of the change for now.