[llvm-objcopy] Add --build-id-link-dir flag

[jakehehrlich]

Summary

This flag does not exist in GNU objcopy but has a major use case. Debugging tools support the .build-id directory structure to find debug binaries. There is no easy way to build this structure up however. One way to do it is by using llvm-readelf and some crazy shell magic. This implements the feature directly. It is most often the case that you'll want to strip a file and send the original to the .build-id directory but if you just want to send a file to the .build-id directory you can copy to /dev/null instead.

I happen to have implemented the "find build id" algorithm like 3 or 4 times now. You can find reference implementations that are known to work here and here.

The .build-id use case documentation can be found here: https://sourceware.org/gdb/onlinedocs/gdb/Separate-Debug-Files.html

In Fuchsia we now have a lot of tools that fetch binaries based on their build id but we do it using this awful non-standard thing we call "ids.txt". This will let us replace it.

[mcgrathr]

llvm/tools/llvm-objcopy/ELF/ELFObjcopy.cpp

139

Is there a check (and test coverage) for the error case where the build ID is less than two bytes?

141

Where's the documentation for the behavior here?
The linked-to gdb documentation describes the .build-id/ directory convention in which it looks for xx/xxx.debug, not xx/xxx.
This is also what lldb implements (it only ever looks for xx/xxx.debug, never xx/xxx).

If the intent here is to publish the full, unstripped input file as the debug file, then it should have the .debug suffix for the gdb convention.
elfutils does look for xx/xxx when looking for an ELF file cold by build ID (which is the use case here, but not something AFAIK that either gdb or lldb supports), and then only looks for xx/xxx.debug if the main file doesn't have debug info.
But that's not the common convention.

When objcopy is being used to strip and produce a separate debug file (with SHT_NOBITS replacements for SHF_ALLOC sections, etc.), then what you want to link as xx/xxx.debug is the separate debug file, not the input file. In that case the best behavior IMHO is to also link xx/xxx to the stripped output file. That's consistent with e.g. Fedora/CentOS/RHEL debuginfo rpms, which include the separate debug file in /usr/lib/debug/.build-id/xx/xxx.debug and a symlink to the installed (stripped) binary as /usr/lib/debug/.build-id/xx/xxx.

142

Is this stray debug output?

145

In POSIX terms, this is usually done by checking for errno==EEXIST rather than another filesystem call to see if that was the reason it failed.

llvm/tools/llvm-objcopy/ELF/Note.h

22 ↗

(On Diff #173532)

static is always superfluous inside an anonymous namespace.
This should be const char *const or constexpr if LLVM allows newer C++.
Except it's unused, so it shouldn't be here at all.

24 ↗

(On Diff #173532)

Actually some systems do use 64-bit fields in ELFCLASS64 SHT_NOTE (or PT_NOTE) data.

91 ↗

(On Diff #173532)

So this tool does not handle cross-endian file access?
And what's wrong with ELFFile::notes() anyway?

[jakehehrlich]

llvm/tools/llvm-objcopy/ELF/ELFObjcopy.cpp

139

This ArrayRef will be empty in that case. There is however no check to make sure that the build ID isn't empty so the [0] above could fail in theory.

141

--only-keep-debug is a nop in llvm-objcopy. There is also no all in one "strip-to-file" option right now. It's certainly doable but it doesn't currently exist. The semantics would be the same weather the input file was copied or we went though the motions.

142

yes.

145

create_hard_link doesn't seem to give any better information that could be used to do this across all systems.

llvm/tools/llvm-objcopy/ELF/Note.h

24 ↗

(On Diff #173532)

llvm doesn't handle them and they're outliers.

91 ↗

(On Diff #173532)

Ah crap...well that *is* an issue that I should fix I suppose.

[alexander-shaposhnikov]

llvm/tools/llvm-objcopy/ELF/Note.h

18 ↗

(On Diff #173532)

i think we should not have anonymous namespaces in headers,
every time this header is included everything will be duplicated.
so i would remove this anonymous namespace, leave only the "public" interface declarations in this header and move the details of implementation into .cpp

22 ↗

(On Diff #173532)

if i remember correctly the existing Visual Studio build bots were not happy with this, though things might have changed.

[alexander-shaposhnikov]

llvm/tools/llvm-objcopy/ELF/Note.h

22 ↗

(On Diff #173532)

(i meant "with this usage of constexpr")

[jakehehrlich]

I started to rewrite it in terms of ELFFile. The main challenge nis dealing
with the Error object that we don't really care about.

[rupprecht]

cc'ing blaike as a local debug expert -- I'm not familiar with systems using --build-id

llvm/test/tools/llvm-objcopy/build-id-link-dir.test

4

Does this also work for in-place invocations?

llvm/tools/llvm-objcopy/ELF/ELFObjcopy.cpp

139

I agree -- this should error (with a test too :) ) if build id is less than 2 bytes

141

nit: comment the true here, e.g. /*LowerCase*/true

143

Why create a hard link instead of copying the file?

llvm/tools/llvm-objcopy/ELF/Note.h

1 ↗

(On Diff #173532)

I don't know if you need this file at all. NoteIterator/NoteRange looks like it's reimplementing the note iterator in libobject
e.g. https://github.com/llvm-mirror/llvm/blob/master/include/llvm/Object/ELF.h#L228

llvm/tools/llvm-objcopy/ObjcopyOpts.td

167

Should this be applied to llvm-strip as well?

[phosek]

llvm/tools/llvm-objcopy/ObjcopyOpts.td

167

Yes, I believe we'll need that as well.

[dblaikie]

@mcgrathr 's comments are probably more relevant than mine - especially those that, by the sounds of it, remove the need for adding a new iterator (but I only found that after I'd already written some of that - and figured the comments might be useful general feedback/'learnings' anyway :) )

llvm/tools/llvm-objcopy/ELF/ELFObjcopy.cpp

130

Prefer = instead of () to make it clear you're not relying on an explicit ctor/conversion?

SmallString<123> Path = Config.BuildIdLinkDir;

llvm/tools/llvm-objcopy/ELF/Note.h

59 ↗

(On Diff #173532)

An Explicit default ctor seems a bit novel - is there anything this is particularly protecting against? (seems like it'd make it invalid to use {} to construct a NoteIterator - is that important?)

Also: Maybe use non-static data member initializers?

const uint64_t *Begin = nullptr;
const uint64_t *End = nullptr;
const Note *Cur = nullptr;

(& then this ctor could be "NoteIterator() = default;")

62 ↗

(On Diff #173532)

Might as well use the init list for Cur too?

66–70 ↗

(On Diff #173532)

Usually folks omit {} on single-line (or single statement) scopes in LLVM code.

(also could use a conditional operator, since each branch is an assignment to the same variable - but that's super subjective, of course :) )

73–74 ↗

(On Diff #173532)

Prefer non-member versions of operator overloads so any implicit conversions apply equally to the LHS and RHS? (they can still be friend definitions defined inline in the class definition here)

& maybe define one in terms of the other for consistency?

83–86 ↗

(On Diff #173532)

Const qualifying value return types is a bit problematic, so far as I recall - doesn't allow the value to be moved from, etc, etc?

90 ↗

(On Diff #173532)

I don't think there's much use of {} ctor invocation in LLVM, so I'd suggest using () for consistency.

90 ↗

(On Diff #173532)

Worth using std::find? Probably not, but figured I'd mention/ask in case.

[dblaikie]

In D54384#1295839, @rupprecht wrote:

cc'ing blaike as a local debug expert -- I'm not familiar with systems using --build-id

Thanks! Had it on my "to look at" list in any case (& have had some discussions with the Fuschia folks about build-id previously) but I really know very little about build-id, where it came from, who uses it/why, how broad the adoption (& thus how reasonable the request for tooling functionality, etc) is. But certainly here to pay attention/provide what feedback & learn what I can from it all.

[jakehehrlich]

So I fixed this issue: https://reviews.llvm.org/D54451

but now I seem to have found a different issue that causes a program failure to happen due to Error being handled improperly. After which we should be using the ELFType iterator implementation.

llvm/test/tools/llvm-objcopy/build-id-link-dir.test

4

It works it's just odd because you modify the the thing you sent to the build id directory.

llvm/tools/llvm-objcopy/ELF/ELFObjcopy.cpp

143

To save space but I guess copying would make the inplace case more sensible. Generally the build system may still need the unstripped file around. For instance, LLD doesn't support linking against binaries stripped with --strip-sections. Because debug binaries total to a rather large amount, not hard-linking would all at once double the size of the build. Currently in Fuchsia, on a lower layer, our build requires ~30gb of space. Of that 10 gb is for unstripped binaries. If we copied we'd suddenly shoot our build size up to 40gb for the smallest part. This isn't a terrible cost mind you but some people have complained about the space requirements. That seems like a high cost to pay to me.

llvm/tools/llvm-objcopy/ELF/Note.h

1 ↗

(On Diff #173532)

Yeah, I'm rewriting everything in terms of that right now. This file will go away.

llvm/tools/llvm-objcopy/ObjcopyOpts.td

167

Sounds good, I'll add it.

[jakehehrlich]

Not technically something we can't work around, the following code solves my problem now that I understand it.

Error Err = Error::success();
if (Err) assert(false); // This only serves to ensure that Err is checked.

This is despite that fact that I check each time before the destructor would be called...I also have to give to the library unchecked as well or else if it assigns an error the program will crash. Will upload change soon.

[jakehehrlich]

llvm/test/tools/llvm-objcopy/build-id-link-dir.test

4

Sorry, this is not right as my tests show. I thought this would be the case but I remembered that FileBuffer atomically renames the file into place which has different semantics. I wrote that comment piror to thinking about that and didn't delete this.

llvm/tools/llvm-objcopy/ELF/ELFObjcopy.cpp

143

So also follow up on this comment as well, most of this is still correct but this behavior is actually exactly as sensible as copying and also saves significant space. Also almost all tools use the rename at the end thing needed to make this true to gain atomicity and avoid causing problems on failure. So I think hard linking is just strictly better.

[rupprecht]

llvm/test/tools/llvm-objcopy/build-id-link-dir.test

6

Is test -e portable (i.e. on windows)? Just ls might be a better option.

llvm/tools/llvm-objcopy/ELF/ELFObjcopy.cpp

123

This should probably error instead of assert, otherwise this will cause weird failures depending on debug vs release builds

I don't really understand why it's needed from this comment either -- can you elaborate the comment?

(Probably the libobject code should just be fixed, but I don't see that as a blocker)

125

The ELF dumper in llvm-readobj only looks for notes in the program headers if it's a core file:
https://github.com/llvm-mirror/llvm/blob/master/tools/llvm-readobj/ELFDumper.cpp#L3883
Honestly I'm not sure why it does it though. Do you need to do similar logic here?

173

BuildIdBytes.getValue().slice(1) instead

576

Is the In here supposed to be Out?

llvm/tools/llvm-objcopy/ObjcopyOpts.td

167

Maybe you just haven't gotten to it yet, but I still don't see the Stripopts.td change in this patch

169

This looks long -- you should be able to format this automatically with git-clang-format --extensions td master

[dblaikie]

llvm/tools/llvm-objcopy/ELF/ELFObjcopy.cpp

155–156

Curious API design - perhaps you wouldn't trip over the Error clearing issues if you used

Expected<ArrayRef<uint8_t>> findBuildID(In);

Errors as out parameters are, tricky.

[alexander-shaposhnikov]

llvm/tools/llvm-objcopy/ELF/ELFObjcopy.cpp

139

as David mentioned below - maybe return Expected<...> instead ? imho that would be more readable, no "Out" parameters, etc ?

[jakehehrlich]

1. Switched to expected
2. Used slice
3. Formatted td file

llvm/tools/llvm-objcopy/ELF/ELFObjcopy.cpp

123

This is very subtle and this assert should never trigger, hence the assert. I changed it to llvm_unreachable instead. Error has state, checked vs unchecked. libObject assigns to the error rather than returning an error, but this is faulty for this reason. You can only construct an error using Error::success() but by default that is unchecked. If you assign to an unchecked Error, your program will assert fail. To avoid this, I just manually check the successful error. This is just an indication that libObject chose a bad way to propagate errors in this case. I can't really avoid it unfortunately.

125

That's not really ideal behavior; ELFDumper should check both. For build ids the correct place to look is always via the program headers because only non-ET_REL binaries have build ids and such binaries will always have a PT_NOTE if they have a build id. This means stripped binaries can (even those stripped with --strip-sections) can be matched up with their build id.

155–156

As I have discovered by trying to deal with the one in libObject you'll see above. I've switched the code to use Expected.

576

No, the build id is unchanged. Thinking about it, we should probably extract the build id once and then pass the build id...not recompute it twice.

llvm/tools/llvm-objcopy/ObjcopyOpts.td

169

woah...I didn't know that was a thing!

[dblaikie]

llvm/tools/llvm-objcopy/ELF/ELFObjcopy.cpp

123

This also probably isn't ideal - ah, here's the tool for this:

http://llvm.org/doxygen/classllvm_1_1ErrorAsOutParameter.html - hopefully that addresses your issues here, specifically "notes" should use an ErrorAsOutParameter to safely clear the error state of the callers new Error object.

137–138

Maybe sink this into the "unwrapOrError" loop (& sink the Error itself into just before the "notes" loop) - shortest scope as needed helps make the code a bit easier to understand/reason about.

[dblaikie]

adding @lhames just as FYI for some stumbling over the Error-as-out-parameter usage.

Ah, found the documentation for it (beyond the basic doxygen): http://llvm.org/docs/ProgrammersManual.html#fallible-constructors

[alexander-shaposhnikov]

I would probably use more detailed help messages for the new options (i.e. what exactly the last two options mean (name's suffix) and how exactly the path is constructed from the build-id),
so that users can understand what's going on without looking at the actual implementation.
other than that - LGTM

[jhenderson]

Sorry for not getting to this earlier. It's been a busy week for me.

llvm/test/tools/llvm-objcopy/build-id-link-dir.test

6

I can't verify for all Windows versions, but on mine, it is part of GnuWin32 tools, which is a prerequisite, I believe.

11

--build-id-link-output= ... looks weird and I don't think should be valid syntax. Could it not just be --build-id-link-output?

If specifying an output is valid, there should be a test for that...

Basically, I'm a little confused as to what is going on here and in the next test.

llvm/test/tools/llvm-objcopy/no-build-id2.test

1

Why is this test a separate test? Perhaps naming it differently would help clarify.

llvm/tools/llvm-objcopy/CopyConfig.cpp

291

clang-format?

llvm/tools/llvm-objcopy/ELF/ELFObjcopy.cpp

125

Why the constexpr here? Pretty sure it's not going to achieve much. Also, I don't think this is the correct case for constants in LLVM - I think it's supposed to be the same as other local variables.

139

Use makeStringError, and llvm::errc::invalid_argument, to avoid the std::make_error_code. It might be nice to also mention the object file name in the error message.

162

It might be nice to add a little bit more to this error message to make it clear what build ID is too small (e.g. "the build ID in file "foo" is smaller than two bytes").

165

/*LowerCase*/

179

Comment the true here too.

182

This should probably mention where it's trying to link to, as well.

llvm/tools/llvm-objcopy/ObjcopyOpts.td

169

Use -> Used
or maybe better yet:
Set directory for --build-id-link-input and --build-id-link-output to <dir>

Using the <dir> meta variable name for the two options in this help text is a bit confusing, given that <dir> is also the name of the meta variable for this option.

[jhenderson]

llvm/test/tools/llvm-objcopy/bad-build-id.test

5

You don't need to capture "T" here, since you don't use it in a later CHECK. You can do {{.*}} instead.

llvm/test/tools/llvm-objcopy/build-id-link-dir.test

11

The weird syntax still exists: trailing '=' looks like something is missing. I'd prefer a solution to it. Either --build-id-link-output should be a valid switch on its own with no argument, as well as --build-id-link-output=<some value> or some other solution si needed.

llvm/tools/llvm-objcopy/ELF/ELFObjcopy.cpp

131–134

clang-format, and unnecessary braces.

139

Oops, yes, createStringError. You don't need the std::make_error_code now.

144

I might be missing something, but I think the parentheses here are all unnecessary.

163

Remove this return and the parentheses, since error() is a no-return function.

P.S. Why do we have separate error and reportError functions?

169–170

Could these two lines be combined to the following?

if (auto EC = sys::fs::create_hard_link(ToLink, Path)) {

[jakehehrlich]

Responded to latest comments.

llvm/test/tools/llvm-objcopy/build-id-link-dir.test

6

I have a windows box with a minimal install for this exact reason. I'll try it later to make sure it works and report back.

11

Setting an output is valid. It's just that stripped binaries traditionally have no extension. an empty "=" is commonly used to pass an empty string and is valid for all "=" arguments across llvm from the argument parser's perspective. This is equivalent to --build-id-link-output ""

The imagined use case for --build-idlink-output that uses a non-empty string would be if you're not stripping, you're adding or removing something and you need ".debug" afterwards. Or using "--only-keep-debug" despite that not having a proper implementation at this exact moment. The example for this would be llvm-objcopy --build-id-link-output=.debug --only-keep-debug <input> <output> there's also this informally proposed idea that has floated around a few people here of adding a third ".dwp" file to the build id directory but I think people wanted to put that feature directly in llvm-dwp. Also in that case you're more or less directly using llvm-objcopy just to do the copy to the build id directory and no other operation.

llvm/test/tools/llvm-objcopy/no-build-id2.test

1

The intent was for one to have notes but no build id entry and the other to just not have notes. Those would need two different yaml files. I guess I forgot to add that. They've been added now.

llvm/tools/llvm-objcopy/ELF/ELFObjcopy.cpp

125

In other code bases I've been working in constexpr has been prefered. As for the case correctness. This is a defined constant in system's <elf.h> but not in llvm's BinaryFormat header for elf. I'm adding it here as it would normally appear in the header. I'll remove the constexpr but I'd like to keep the name.

163

1. Because I copied the error handeling functions from other places in the very first commit and we've just kept it so far
2. They all do slightly different things. Though I think I'm not consistent about when I use the Error form of reportError and when I manually use error() with an Error

169–170

yep. Thanks!

179

I realized it wasn't needed and dropped in instead. I commented the other one as requested however.

[jhenderson]

LGTM, with a couple of minor comments.

llvm/test/tools/llvm-objcopy/build-id-link-dir.test

11

Ah, okay that makes more sense now. Thanks. "" is a bit clearer to me, I guess.

llvm/test/tools/llvm-objcopy/no-build-id2.test

1

Okay. Maybe rename "no-build-id2.test" to "no-build-id-no-notes.test"? Not too bothered though. Alternatively, a comment in the two files explaining the difference.

llvm/tools/llvm-objcopy/ELF/ELFObjcopy.cpp

125

Could we not just add it to the header?

[jakehehrlich]

Aight, uploading the final patch so people can see exactly what I'm going to push.

llvm/test/tools/llvm-objcopy/bad-build-id.test

+# RUN: yaml2obj %s > %t
+# RUN: not llvm-objcopy --build-id-link-dir=%t-dir --build-id-link-input=.debug %t 2>&1 >/dev/null | FileCheck %s
+
+# CHECK: build ID size is too small
+
+--- !ELF
+FileHeader:
+  Class:           ELFCLASS64
+  Data:            ELFDATA2LSB
+  Type:            ET_EXEC
+  Machine:         EM_X86_64
+Sections:
+  - Name:            .note.gnu.build-id
+    Type:            SHT_NOTE
+    Flags:           [ SHF_ALLOC ]
+    Content:         040000000100000003000000474E55004F000000
+ProgramHeaders:
+  - Type: PT_NOTE
+    Flags: [ PF_R ]
+    Sections:
+      - Section: .note.gnu.build-id

llvm/test/tools/llvm-objcopy/build-id-link-dir.test

+# RUN: yaml2obj %s > %t
+# RUN: mkdir -p %t-dir
+
+# RUN: llvm-objcopy --build-id-link-dir=%t-dir %t %t2
+# RUN: not test -e %t-dir/4f/cb712aa6387724a9f465a32cd8c14b
+
+# RUN: llvm-objcopy --build-id-link-dir=%t-dir --build-id-link-input=.debug %t %t3
+# RUN: cmp %t-dir/4f/cb712aa6387724a9f465a32cd8c14b.debug %t
+# RUN: rm %t-dir/4f/cb712aa6387724a9f465a32cd8c14b.debug
+
+# RUN: llvm-objcopy --build-id-link-dir=%t-dir --build-id-link-output= --strip-sections %t %t4
+# RUN: cmp %t-dir/4f/cb712aa6387724a9f465a32cd8c14b %t4
+# RUN: rm %t-dir/4f/cb712aa6387724a9f465a32cd8c14b
+
+# Linking the output of an inplace argument means that the file in the build-id
+# directory will be hard-linked to the resulting file.
+# RUN: cp %t %t5
+# RUN: llvm-objcopy --build-id-link-dir=%t-dir --build-id-link-output= --strip-sections %t5
+# RUN: cmp %t-dir/4f/cb712aa6387724a9f465a32cd8c14b %t5
+# RUN: rm %t-dir/4f/cb712aa6387724a9f465a32cd8c14b
+
+# Linking the input of an inplace argument means that the file in the build-id
+# directory will be hard-linked to the original file.
+# RUN: cp %t %t6
+# RUN: llvm-objcopy --build-id-link-dir=%t-dir --build-id-link-input=.debug --strip-sections %t6
+# RUN: cmp %t-dir/4f/cb712aa6387724a9f465a32cd8c14b.debug %t
+# RUN: rm %t-dir/4f/cb712aa6387724a9f465a32cd8c14b.debug
+
+# You can use both at once.
+# RUN: llvm-objcopy --build-id-link-dir=%t-dir --build-id-link-output= --build-id-link-input=.debug --strip-sections %t %t7
+# RUN: cmp %t-dir/4f/cb712aa6387724a9f465a32cd8c14b.debug %t
+# RUN: cmp %t-dir/4f/cb712aa6387724a9f465a32cd8c14b %t7
+# RUN: rm %t-dir/4f/cb712aa6387724a9f465a32cd8c14b.debug
+# RUN: rm %t-dir/4f/cb712aa6387724a9f465a32cd8c14b
+
+--- !ELF
+FileHeader:
+  Class:           ELFCLASS64
+  Data:            ELFDATA2LSB
+  Type:            ET_EXEC
+  Machine:         EM_X86_64
+Sections:
+  - Name:            .note.gnu.build-id
+    Type:            SHT_NOTE
+    Flags:           [ SHF_ALLOC ]
+    Content:         040000001000000003000000474E55004FCB712AA6387724A9F465A32CD8C14B
+ProgramHeaders:
+  - Type: PT_NOTE
+    Flags: [ PF_R ]
+    Sections:
+      - Section: .note.gnu.build-id

llvm/test/tools/llvm-objcopy/no-build-id.test

+# RUN: yaml2obj %s > %t
+# RUN: not llvm-objcopy --build-id-link-dir=%t-dir --build-id-link-input=.debug %t 2>&1 >/dev/null | FileCheck %s
+
+# CHECK: Could not find build ID.
+
+--- !ELF
+FileHeader:
+  Class:           ELFCLASS64
+  Data:            ELFDATA2LSB
+  Type:            ET_EXEC
+  Machine:         EM_X86_64
+Sections:

llvm/test/tools/llvm-objcopy/no-build-id2.test

+# RUN: yaml2obj %s > %t
+# RUN: not llvm-objcopy --build-id-link-dir=%t-dir --build-id-link-input=.debug %t 2>&1 >/dev/null | FileCheck %s
+
+# CHECK: Could not find build ID.
+
+--- !ELF
+FileHeader:
+  Class:           ELFCLASS64
+  Data:            ELFDATA2LSB
+  Type:            ET_EXEC
+  Machine:         EM_X86_64

llvm/tools/llvm-objcopy/CopyConfig.h

struct CopyConfig {
   StringRef OutputFilename;
   StringRef OutputFormat;
 
   // Only applicable for --input-format=Binary
   MachineInfo BinaryArch;
 
   // Advanced options
   StringRef AddGnuDebugLink;
+  StringRef BuildIdLinkDir;
+  Optional<StringRef> BuildIdLinkInput;
+  Optional<StringRef> BuildIdLinkOutput;
   StringRef SplitDWO;
   StringRef SymbolsPrefix;
 
   // Repeated options
   std::vector<StringRef> AddSection;
   std::vector<StringRef> DumpSection;
   std::vector<StringRef> Keep;
   std::vector<StringRef> OnlyKeep;

llvm/tools/llvm-objcopy/CopyConfig.cpp

if (Arg->getOption().getID() == OBJCOPY_compress_debug_sections_eq) {
       if (Config.CompressionType == DebugCompressionType::None)
         error("Invalid or unsupported --compress-debug-sections format: " +
               InputArgs.getLastArgValue(OBJCOPY_compress_debug_sections_eq));
       if (!zlib::isAvailable())
         error("LLVM was not compiled with LLVM_ENABLE_ZLIB: can not compress.");
     }
   }
 
-  Config.SplitDWO = InputArgs.getLastArgValue(OBJCOPY_split_dwo);
   Config.AddGnuDebugLink = InputArgs.getLastArgValue(OBJCOPY_add_gnu_debuglink);
+  Config.BuildIdLinkDir = InputArgs.getLastArgValue(OBJCOPY_build_id_link_dir);
+  if (InputArgs.hasArg(OBJCOPY_build_id_link_input))
+    Config.BuildIdLinkInput = InputArgs.getLastArgValue(OBJCOPY_build_id_link_input);
+  if (InputArgs.hasArg(OBJCOPY_build_id_link_output))
+    Config.BuildIdLinkOutput = InputArgs.getLastArgValue(OBJCOPY_build_id_link_output);
+  Config.SplitDWO = InputArgs.getLastArgValue(OBJCOPY_split_dwo);
   Config.SymbolsPrefix = InputArgs.getLastArgValue(OBJCOPY_prefix_symbols);
 
   for (auto Arg : InputArgs.filtered(OBJCOPY_redefine_symbol)) {
     if (!StringRef(Arg->getValue()).contains('='))
       error("Bad format for --redefine-sym");
     auto Old2New = StringRef(Arg->getValue()).split('=');
     if (!Config.SymbolsToRename.insert(Old2New).second)
       error("Multiple redefinition of symbol " + Old2New.first);

llvm/tools/llvm-objcopy/ELF/ELFObjcopy.cpp

 //===- ELFObjcopy.cpp -----------------------------------------------------===//
 //
 //                      The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 
 #include "ELFObjcopy.h"
 #include "Buffer.h"
 #include "CopyConfig.h"
-#include "llvm-objcopy.h"
 #include "Object.h"
+#include "llvm-objcopy.h"
 
 #include "llvm/ADT/BitmaskEnum.h"
 #include "llvm/ADT/Optional.h"
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/ADT/Twine.h"
 #include "llvm/BinaryFormat/ELF.h"
return llvm::make_unique<ELFWriter<ELF32BE>>(Obj, Buf,
                                                  !Config.StripSections);
   case ELFT_ELF64BE:
     return llvm::make_unique<ELFWriter<ELF64BE>>(Obj, Buf,
                                                  !Config.StripSections);
   }
   llvm_unreachable("Invalid output format");
 }
 
+template <class ELFT>
+static Expected<ArrayRef<uint8_t>>
+findBuildID(const object::ELFFile<ELFT> &In) {
+  Error Err = Error::success();
+  // Make sure this is unchecked and tangentially check that it's not an error.
+  if (Err)
+    llvm_unreachable("This error should have been a success");
+  constexpr const char *ELF_NOTE_GNU = "GNU";
+  for (const auto &Phdr : unwrapOrError(In.program_headers())) {
+    if (Phdr.p_type != PT_NOTE)
+      continue;
+    for (const auto &Note : In.notes(Phdr, Err)) {
+      if (Err)
+        return std::move(Err);
+      if (Note.getType() == NT_GNU_BUILD_ID && Note.getName() == ELF_NOTE_GNU) {
+        return Note.getDesc();
+      }
+    }
+  }
+  if (Err)
+    return std::move(Err);
+  return make_error<StringError>(
+      "Could not find build ID.",
+      std::make_error_code(std::errc::invalid_argument));
+}
+
+static Expected<ArrayRef<uint8_t>>
+findBuildID(const object::ELFObjectFileBase &In) {
+  if (auto *O = dyn_cast<ELFObjectFile<ELF32LE>>(&In)) {
+    return findBuildID(*O->getELFFile());
+  } else if (auto *O = dyn_cast<ELFObjectFile<ELF64LE>>(&In)) {
+    return findBuildID(*O->getELFFile());
+  } else if (auto *O = dyn_cast<ELFObjectFile<ELF32BE>>(&In)) {
+    return findBuildID(*O->getELFFile());
+  } else if (auto *O = dyn_cast<ELFObjectFile<ELF64BE>>(&In)) {
+    return findBuildID(*O->getELFFile());
+  }
+  llvm_unreachable("Bad file format");
+}
+
+static void linkToBuildIdDir(StringRef Dir, StringRef ToLink, StringRef Suffix,
+                             const object::ELFObjectFileBase &In) {
+  ArrayRef<uint8_t> BuildIdBytes = unwrapOrError(findBuildID(In));
+  if (BuildIdBytes.size() < 2)
+    error("build ID size is too small");
+
+  SmallString<128> Path = Dir;
+  sys::path::append(Path, llvm::toHex(BuildIdBytes[0], true));
+  if (auto EC = sys::fs::create_directories(Path)) {
+    error("cannot create build ID link directory " + Path + ": " +
+          EC.message());
+    return;
+  }
+
+  sys::path::append(Path,
+                    llvm::toHex(BuildIdBytes.slice(1), /*LowerCase*/ true));
+  Path += Suffix;
+  auto EC = sys::fs::create_hard_link(ToLink, Path);
+  if (EC) {
+    // Hard linking failed, try to remove the file first if it exists.
+    if (sys::fs::exists(Path))
+      sys::fs::remove(Path, true);
+    EC = sys::fs::create_hard_link(ToLink, Path);
+    if (EC)
+      error("cannot link " + Path + ": " + EC.message());
+  }
+}
+
 static void splitDWOToFile(const CopyConfig &Config, const Reader &Reader,
                            StringRef File, ElfType OutputElfType) {
   auto DWOFile = Reader.create();
   DWOFile->removeSections(
       [&](const SectionBase &Sec) { return onlyKeepDWOPred(*DWOFile, Sec); });
   FileBuffer FB(File);
   auto Writer = createWriter(Config, *DWOFile, FB, OutputElfType);
   Writer->finalize();
void executeObjcopyOnRawBinary(const CopyConfig &Config, MemoryBuffer &In,
   Writer->write();
 }
 
 void executeObjcopyOnBinary(const CopyConfig &Config,
                             object::ELFObjectFileBase &In, Buffer &Out) {
   ELFReader Reader(&In);
   std::unique_ptr<Object> Obj = Reader.create();
   const ElfType OutputElfType = getOutputElfType(In);
+  if (!Config.BuildIdLinkDir.empty() && Config.BuildIdLinkInput) {
+    linkToBuildIdDir(Config.BuildIdLinkDir, Config.InputFilename,
+                     Config.BuildIdLinkInput.getValue(), In);
+  }
   handleArgs(Config, *Obj, Reader, OutputElfType);
   std::unique_ptr<Writer> Writer =
       createWriter(Config, *Obj, Out, OutputElfType);
   Writer->finalize();
   Writer->write();
+  if (!Config.BuildIdLinkDir.empty() && Config.BuildIdLinkOutput) {
+    linkToBuildIdDir(Config.BuildIdLinkDir, Config.OutputFilename,
+                     Config.BuildIdLinkOutput.getValue(), In);
+  }
 }
 
 } // end namespace elf
 } // end namespace objcopy
 } // end namespace llvm

llvm/tools/llvm-objcopy/ELF/Object.h

bool operator()(const SectionBase *Lhs, const SectionBase *Rhs) const {
       // address of the actully stored section.
       if (Lhs->OriginalOffset == Rhs->OriginalOffset)
         return Lhs < Rhs;
       return Lhs->OriginalOffset < Rhs->OriginalOffset;
     }
   };
 
   std::set<const SectionBase *, SectionCompare> Sections;
-  ArrayRef<uint8_t> Contents;
 
 public:
   uint64_t Align;
   uint64_t FileSize;
   uint32_t Flags;
   uint32_t Index;
   uint64_t MemSize;
   uint64_t Offset;
   uint64_t PAddr;
   uint64_t Type;
   uint64_t VAddr;
 
   uint64_t OriginalOffset;
   Segment *ParentSegment = nullptr;
+  ArrayRef<uint8_t> Contents;
 
   explicit Segment(ArrayRef<uint8_t> Data) : Contents(Data) {}
   Segment() {}
 
   const SectionBase *firstSection() const {
     if (!Sections.empty())
       return *Sections.begin();
     return nullptr;

llvm/tools/llvm-objcopy/ObjcopyOpts.td

: Eq<"dump-section",
          "Dump contents of section named <section> into file <file>">,
       MetaVarName<"section=file">;
 defm prefix_symbols
     : Eq<"prefix-symbols", "Add <prefix> to the start of every symbol name">,
       MetaVarName<"prefix">;
 
 def version : Flag<["-", "--"], "version">,
               HelpText<"Print the version and exit.">;
+defm build_id_link_dir
+    : Eq<"build-id-link-dir",
+         "Use as <dir> by --build-id-link-input and --build-id-link-output">,
+      MetaVarName<"dir">;
+defm build_id_link_input
+    : Eq<"build-id-link-input", "Hard-link the input to <dir>/xx/xxx<suffix> "
+                                "name derived from hex build ID">,
+      MetaVarName<"suffix">;
+defm build_id_link_output
+    : Eq<"build-id-link-output", "Hard-link the output to <dir>/xx/xxx<suffix> "
+                                 "name derived from hex build ID">,
+      MetaVarName<"suffix">;