[AIX] Support of Big archive (read)

[EGuesnet]

Summary

Add support of AIX Big Archive. Read only.

[jhenderson]

llvm/include/llvm/Object/Archive.h

373–375

Do you really need this length? Why is it uint32_t?

378

This is completely unthread safe - if you have a tool parsing two archives simultaneously, this will fail horribly. Do you really need it though? See my comments further down.

From reading the spec, the archive uses offsets everywhere and it should be simple to use the values stored in the member headers.

llvm/lib/Object/Archive.cpp

485

Don't add this comment. We can see that these are child constructors by the function signature.

490

Use std::make_unique to create a unique pointer without needing to explicitly mention new. Same everywhere else you've done this.

500

Don't add a blank line at the function start.

649

It seems to me like the logic you've added to this function is a little odd. I think you may be trying too hard to match the existing logic, when it doesn't really make sense. In traditional UNIX ar format, each child is immediately followed by the next one (barring a possible even-alignment padding byte), stopping when you get to the end. My reading of the Big Archive spec is that the next child's offset is defined by the ar_nxtmem member of its header. You should just be using that and fl_lstmoff to identify if the current child is the last one or not. You can then check to see if the child goes past the buffer end, as per the existing check in this file.

[EGuesnet]

[AIX] Support of Big archive (read)

Add support of AIX Big Archive. Read only.

[EGuesnet]

You should get some people who are more familiar with XCOFF/AIX to help review correctness. There has been a lot of development for this in other tools (e.g. llvm-readobj).

The IBM team that has sent PR about llvm-readobj is aware of this PR. I will recontact them.

I think I have respond to most of your comment. The main issue is the static variable. I need to know the current location to distinguish Fix length header and Object headers, and I have not found cleaner solution.
Note that it has been checked about multiple read: Rust loads libLLVM.so once, and can read multiple archive. It works, as the static variable is initialized correctly before being used. But it is sequential read.

llvm/include/llvm/Object/Archive.h

373–375

Maximal offset (in fixed length header) is stored as 20 bytes as decimal digit. So, maximal length is approximately 10^10. It is the total length of the archive. So, it must be uint64-t, not uint32_t.

378

Main trouble is with first read, to distinguish Fixed length header and other one.
We must know the current location, and modify it during read. This cannot be done in Parent (because of const) or Header (not access when it is required). Static variable in archive is the least bad solution I have found.

llvm/lib/Object/Archive.cpp

649

I think you may be trying too hard to match the existing logic

Right.

My reading of the Big Archive spec is that the next child's offset is defined by the ar_nxtmem member of its header. You should just be using that and fl_lstmoff to identify if the current child is the last one or not. You can then check to see if the child goes past the buffer end, as per the existing check in this file.

But using offset needs to know the current location. You say in another comment this his highly not thread-safe and might be avoid.
Moreover, if I use fl_lstmoff instead of Length, I must have an access to Fixed length header. It is currently read and forbidden as length of object in archive in the only information we need.

1017

I have removed code related to symbol read.

llvm/lib/Object/ArchiveWriter.cpp

141

It can be reached by external tools (not llvm-ar), so I modifie it to use report_fatal_error().

llvm/test/Object/archive-big-read.test

2

I think it is OK now.
I have first done the test inspiring from other one, and they do not use theses convention and flag.

4

Done. See archive-big-extract.test.

[EGuesnet]

Use of clang-format.

[jhenderson]

I need to know the current location to distinguish Fix length header and Object headers, and I have not found cleaner solution.

I still don't understand why you need to know it to make this distinction. The two header types are completely different things as far as I understand the spec, used in different contexts. Could you read the Fixed Length header as the first part of reading the archive, and the object headers when reading individual children, then store them as members of the respective classes?

Note that it has been checked about multiple read: Rust loads libLLVM.so once, and can read multiple archive. It works, as the static variable is initialized correctly before being used. But it is sequential read.

Just because this works in one single case does not make this an appropriate design - in fact you haven't even tested the case where the issue is, as I bet that the thing you ran read the archives sequentially, not in parallel. What if I wrote a tool that wanted to read multiple Big-format archives in parallel, with one thread per archive? What would happen to the static variable on each individual thread?

Pseudo-code example reading algorithm that doesn't have the issues mentioned:

class Archive {
    FixedLengthHeader Header;
    std::vector<Child> Children;
    ArrayRef<uint8_t> Data;

    Archive(ArrayRef<uint8_t> Data) : Data(Data) {
      Header = reinterpret_cast<FixedLengthHeader> (Data.data());
      uint64_t Offset = Header->FirstChildOffset;
      // Check if archive is empty.
      Child C(Offset, this);
      while (C != nullptr) {
        Children.push_back(C);
        C = C.nextChild();
      }
    }
};

class Child {
    ObjectHeader Header;
    uint8_t *Payload;
    Archive &Parent;
    Child(uint64_t Offset, Archive &Archive) {
      // Check Archive data is large enough for offset + header size + payload size.
      Parent = Archive;
      Header = reinterpret_cast<ObjectHeader>(Parent->Data.data() + Offset);
      Payload = Parent->Data.data() + Offset + sizeof(Header);
    }
    Child *nextChild() {
      if (Header->nextOffset == 0)
        return nullptr;
      return Child(Header->nextOffset, Parent);
    }
};

llvm/lib/Object/Archive.cpp

649

It is currently read and forbidden as length of object in archive in the only information we need.

I don't understand this comment at all. Do you mean forgotten, not forbidden? Perhaps you could change how things are read and stored?

llvm/test/Object/archive-big-read.test

2

Older tests use older styles. We have developed better styles in the meantime which are preferred for newer tests.

[EGuesnet]

Update D100651.

[EGuesnet]

Hi @jhenderson,
I cannot follow exactly what you propose, because llvm-ar (and maybe other tools) uses directly child_iterator. As this object does not have access to Archive (=Parent), it seems difficult to transfert Child vector of Archive to child_iterator. Moreover, I want to avoid to un-const Parent in Child.

Using the same idea, but not the same exact way, I propose:

• Move FixedLengthHeader on Archive, and use it for Big Archive only.
• Deal with first read (Fixed Length Header) in Archive creation, so any Child after do not care about it (one of the reason of static variable in previous version).
• To distinguish Object and Member Table, I cannot use Next Member in Object Header, because it points to Member Table in the last Object; I cannot use Offset to Member Table of Fixed Length Header, because I do not memorize the current position. So, I use a trick: an Object must have a name (adding empty-file to an archive is non valid), and Member Table Header does not have name, so name field is "`\n" (terminator).
• The child_begin_bigarchive() function is equivalent to child_begin(), but it deals correctly with Fixed Length Header.

getSize() and Big Archive Header initialiser are simplified, differences between usual Archive and Big Archive is deal during Child and Archive initialisation (so, correctly separated).

Free List is not yet supported: if you remove object in Archive, it will become unreadable by LLVM. Free List is complex, and not fully documented. I use report_fatal_error() if Free List is present in Archive (so, not 0), or if it is impossible to read it.

[Esme]

llvm/test/Object/archive-big-extract.test

2

Can you please add the command of creating the ar file to the comment? Base on this patch, I am trying to enable tools (e.g. llvm-readobj) for ar format, but I am getting an error from BigArchiveMemberHeader::getSize() when applying this patch.

# Create the ar file under AIX
$ ar -v -q 1.a 1.o
$ llvm-ar p 1.a evenlen
llvm-ar: error: unable to load '2.a': truncated or malformed archive (characters in size field in archive header are not all decimal numbers: '\000\000\000\000\000\000\000\000\000\0005892' for archive member header at offset 128)

[EGuesnet]

llvm/test/Object/archive-big-extract.test

2

Hi @Esme ,
I just procede as:

$ # On AIX
$ echo evenlen > evenlen
$ ar -v -q 1.a evenlen
$ # On any server
$ llvm-ar p 1.a evenlen
evenlen

As your error message is related to 2.a, I think there is something wrong to reproduce your error.

[Esme]

llvm/lib/Object/Archive.cpp

1012

It's not correct to calculate the offset of the first archive member by Data.getBufferStart() + strlen(Magic) + sizeof(Archive::ArFixLenHdrType);, please use the value of ArFixLenHdr->FirstArOffset.
You can easily reproduce the bug if you test a archive file which contains a object file member like the comment I added before:

$ xlc 1.c -o 1.o
$ ar -v -q 1.a 1.o
$ llvm-ar tv 1.a
llvm-ar: error: unable to load '1.a': truncated or malformed archive (characters in size field in archive header are not all decimal numbers: '\000\000\000\000\000\000\000\000\000\0005892' for archive member header at offset 128)

The correct offset for this case should be 138 instead of 128.

[EGuesnet]

llvm/lib/Object/Archive.cpp

1012

First, LLVM is no more a priority of our team, so I cannot spend time to this PR.
Second, I am not able to reproduce. Please give me content of 1.c, and content of 1.a.

// small.c
int add_two (int a) {
        return a + 2;
}

$ xlc -v
C for AIX Compiler, Version 5

OK, really old, but we don't care for archive.

$ xlc small.c  -o small.o -c
$ ar -v -q small.a small.o

$ xxd small.a

00000000: 3c62 6967 6166 3e0a 3130 3730 2020 2020  <bigaf>.1070
00000010: 2020 2020 2020 2020 2020 2020 3132 3332              1232
00000020: 2020 2020 2020 2020 2020 2020 2020 2020
00000030: 3020 2020 2020 2020 2020 2020 2020 2020  0
00000040: 2020 2020 3132 3820 2020 2020 2020 2020      128
00000050: 2020 2020 2020 2020 3132 3820 2020 2020          128
00000060: 2020 2020 2020 2020 2020 2020 3020 2020              0
00000070: 2020 2020 2020 2020 2020 2020 2020 2020
// End of Fixed-Length Header: size is 128
00000080: 3832 3020 2020 2020 2020 2020 2020 2020  820
00000090: 2020 2020 3130 3730 2020 2020 2020 2020      1070
000000a0: 2020 2020 2020 2020 3020 2020 2020 2020          0
000000b0: 2020 2020 2020 2020 2020 2020 3136 3331              1631
000000c0: 3630 3435 3039 2020 3020 2020 2020 2020  604509  0
000000d0: 2020 2020 3020 2020 2020 2020 2020 2020      0
000000e0: 3634 3420 2020 2020 2020 2020 3720 2020  644         7
[...]

$ llvm-ar --version
LLVM version 13.0.0

$ llvm-ar t small.a
small.o # OK

Third, according documentation https://www.ibm.com/docs/en/aix/7.2?topic=formats-ar-file-format-big , Fixed-Length Header must have a size of 128. 138 is not valide.

[Esme]

llvm/lib/Object/Archive.cpp

1012

Thanks for your reply!

1. If you are no longer following up on this patch, I'd be happy to continue your work. How do you think?

2. In my previous comment, a binary member is added to the archive, while in your case, there is an object member (compiled with -c option) in archive.

You can also reproduce the issue if you add a dynamic lib to an archive.

$ cat 1.c
int main() {
return 1;
}
$ xlc 1.c -qmkshrobj -o libt.so
$ ar -v -q 1.a libt.so
$ xxd 1.a 
00000000: 3c62 6967 6166 3e0a 3134 3532 2020 2020  <bigaf>.1452    
00000010: 2020 2020 2020 2020 2020 2020 3136 3134              1614
00000020: 2020 2020 2020 2020 2020 2020 2020 2020                  
00000030: 3020 2020 2020 2020 2020 2020 2020 2020  0               
00000040: 2020 2020 3133 3420 2020 2020 2020 2020      134         
00000050: 2020 2020 2020 2020 3133 3420 2020 2020          134     
00000060: 2020 2020 2020 2020 2020 2020 3020 2020              0   
00000070: 2020 2020 2020 2020 2020 2020 2020 2020                  
00000080: 0000 0000 0000 3131 3935 2020 2020 2020  ......1195    
// There are some padding between the Fixed-Length Header and the First member.  
00000090: 2020 2020 2020 2020 2020 3134 3532 2020            1452  
000000a0: 2020 2020 2020 2020 2020 2020 2020 3020                0 
000000b0: 2020 2020 2020 2020 2020 2020 2020 2020

3. It's correct that Fixed-Length Header must have a size of 128. 138 is not valide. As above, there may be some padding between the Fixed-Length Header and the first member. So I think the value of ArFixLenHdr->FirstArOffset is the exact offset to the first member.

[EGuesnet]

llvm/lib/Object/Archive.cpp

1012

Using ArFixLenHdr->FirstArOffset does not work.

truncated or malformed archive (characters in name length field in archive header are not all decimal numbers: '' for archive member header at offset 68)

I do not know where this "68" is from.

In my opinion, this PR is complex enough. It might be accepted with as few changes as possible. So, without freelist, undocumented features... In a second time, you can create a new PR, to extend support of Big Archive, with new tests.

[Esme]

llvm/lib/Object/Archive.cpp

1012

OK, thanks!
Do you get the FirstArOffset value via StringRef(ArFixLenHdr->FirstArOffset, sizeof(ArFixLenHdr->FirstArOffset)).rtrim(' ').getAsInteger(10, Size) ?

[EGuesnet]

llvm/lib/Object/Archive.cpp

1012

My mistake, you are right. I take FirstArOffset directly...
But in order to treat the padded case, you must add test, with binary file. LLVM community want to avoid adding binary to test. So, I think again a separate PR after this one is preferable.

[DiggerLin]

llvm/lib/Object/Archive.cpp

104

There are common source code in the BigArchiveMemberHeader::BigArchiveMemberHeader and ArchiveMemberHeader::ArchiveMemberHeader , we can put the common code in the constructor of AbstractArchiveMemberHeader

[DiggerLin]

llvm/include/llvm/Object/Archive.h

345

I do not think we need a special child_begin_bigarchive() API for aix big ar here,
we can use the same API child_begin() , there is a member
Format in the Archive, we use have different implement of child_begin() base on the value of "Format" , if the
Format is K_AIXBIG(), we have different implement of child_begin().

[DiggerLin]

gentle ping, we need the patch, can you help to speed it up ? @EGuesnet

[EGuesnet]

Hi @DiggerLin,

I do not have time to improve this PR before begin of November. We have other priorities right now.
This PR has nearly 6 months, so, it can wait one more.

[DiggerLin]

llvm/include/llvm/Object/Archive.h

104–155

both the ArchiveMemberHeader and the BigArchiveMemberHeader has the "Archive const *Parent;" , can the member put in the AbstractArchiveMemberHeader ?

llvm/lib/Object/Archive.cpp

62

if the RawHeaderPtr == nullptr is true , the value of ArMemHdr will be random here.

108

same problem as above.

175

the variable name "end" can not express the mean, maybe be "NameSize" or "NameLen" ?

337–392

there are several place use the almost the same code from line 308 to 320 . can we change these lines into a helper function?

564

change "Name.size() + Name.size() % 2" to ((Name.size()+1) >>1 ) << 1 ?

573

the Name.startswith("#1/")) is not for the Archive::K_AIXBIG. what happen if the AIXBIG archive file which has member name begin with "#/" here ?

[DiggerLin]

llvm/lib/Object/Archive.cpp

281

it looks the parameter Size not be used here?

[DiggerLin]

llvm/include/llvm/Object/Archive.h

345

also using a separate interface child_begin_bigarchive , you can not use the API

iterator_range<child_iterator> children(Error &Err,
                                        bool SkipInternal = true) const {
  return make_range(child_begin(Err, SkipInternal), child_end());

[DiggerLin]

the function file_magic llvm::identify_magic(StringRef Magic) should be modified to add big ar identification. otherwise object::createBinary will be error for big ar.

[DiggerLin]

llvm/include/llvm/Object/Archive.h

345

or you can call child_begin_bigarchive() in child_begin(Error &Err, bool SkipInternal = true) when Format is K_AIXBIG()

[DiggerLin]

llvm/lib/Object/Archive.cpp

337–392

and I just wonder why you need the code

std::string Buf;
    raw_string_ostream OS(Buf);
    OS.write_escaped(
        StringRef(ArMemHdr->Size, sizeof(ArMemHdr->Size)).rtrim(" "));
    OS.flush();

to convert to std::string.

for the StringRef , there is a function str() to convert to std::string, why not use str() of StringRef ?

[DiggerLin]

can you address the comment ,update the patch? we will approve it and We will create a following patch for the "Support of Big archive" @EGuesnet

[jsji]

In my opinion, this PR is complex enough. It might be accepted with as few changes as possible. So, without freelist, undocumented features... In a second time, you can create a new PR, to extend support of Big Archive, with new tests.

As @EGuesnet mentioned above, I think he won't have time to address the comments in time.
However, we do depend on this patch for AIX in near future,
so I would suggest that we *accept* this patch *as it is* for now, and commit it as @EGuesnet suggested.

And @DiggerLin will post follow up patches immediately.

Is that OK for reviewers? @jhenderson @rupprecht @Esme ?

[jhenderson]

In D100651#3058324, @jsji wrote:

In my opinion, this PR is complex enough. It might be accepted with as few changes as possible. So, without freelist, undocumented features... In a second time, you can create a new PR, to extend support of Big Archive, with new tests.

As @EGuesnet mentioned above, I think he won't have time to address the comments in time.
However, we do depend on this patch for AIX in near future,
so I would suggest that we *accept* this patch *as it is* for now, and commit it as @EGuesnet suggested.

And @DiggerLin will post follow up patches immediately.

Is that OK for reviewers? @jhenderson @rupprecht @Esme ?

This patch is not in a suitable state ready for commit. I've highlighted a high number of issues, and I think there are some fundamental issues with the approach, which need addressing before it is acceptable to land this in LLVM. I haven't finished reviewing the whole patch yet either. I'm sorry if you need the patch in the near future. I can only suggest that you try implementing something yourself, perhaps by adopting and improving this patch, or starting an alternative version (inspired by this patch, and with credit given to @EGuesnet for their initial work, in the commit message).

llvm/include/llvm/Object/Archive.h

40–42

This comment is superfluous. clone always means this.

51–56

getOffset isn't named getRawOffset, so doesn't look like it belong in this block.

getRawName on the other hand looks like it does belong here.

Please make sure all comments follow the existing style. At the moment, comments are doxygen style (see e.g. getName). LLVM style also uses trailing full stops on comments - please add them where they are missing. That being said, don't bother with comments that provide no useful benefit beyond the method names. For example, you do not need this comment or the "Non-raw getters" comment: what benefit do they provide?

109

Please order const according to standard LLVM style, i.e. const T *.

If you're changing a function anyway, it makes sense to fix the style.

111

Please separate unrelated method declarations, or methods with non-single-line definitions with a blank line.

181

Do we actually need all these new constructors and assignment operators? Are they used?

183

Under what circumstances can C.Header be a nullptr?

186

Nit: add new line before this method.

187

Parent is a raw pointer, there's no need to std::move it.

374

Nit: please use // comments in this struct, as per the rest of the code.

As I understand it, this is a specific BigArchive class. Please rename it to be clear, e.g. BigArchiveFixLenHdr.

382

I think the presence of this member shows that the design is not right.

BigArchive is a kind of Archive. The members specific to BigArchive types (as opposed to e.g. traditional BSD or GNU archives) should not be present in a type that should be generic.

I think rather than, or possibly as well as, having this member, you really want to have concrete Archive subclasses that implement the relevant logic, and make Archive an abstract class.

llvm/lib/Object/Archive.cpp

46

I don't like this define. It is not guaranteed that all future archive magic is 8 characters long. I don't think it would be necessary if we made the Archive class a class hierarchy as suggested above.

I'm also slightly surprised that BigMagic doesn't start with !, but I assume that is correct and intentional.

66

Why has this code moved?

129

Terminator characters are cosmetic only, really, in regular archives too, but we still check them.

171–172

This comment has several typos in, and I'm not really sure what it is actually trying to say.

317–319

I'd prefer that this calculation be factored out into at least a variable, as I don't follow what it represents.

340

This pattern is a duplicate of most of the earlier expression. Put it in a variable rather than duplicating code.

430

getRawName appears not to trim the trailing spaces, but getRawUID does. This seems like a logical inconsistency that will likely lead to bugs.

506

These if/else checks clearly indicate that Child should really be BigArchiveChild and something else, sharing a common Child abstract base class.

660–665

Revert this change.

700–702

This looks like an unrelated formatting change.

[jsji]

I can only suggest that you try implementing something yourself, perhaps by adopting and improving this patch, or starting an alternative version (inspired by this patch, and with credit given to @EGuesnet for their initial work, in the commit message).

Thanks @jhenderson ! That would work too . @DiggerLin Let us do it -- you can create another patch based on this and give credit to @EGuesnet if he can NOT address the comment in time.

[DiggerLin]

In D100651#3058580, @jsji wrote:

I can only suggest that you try implementing something yourself, perhaps by adopting and improving this patch, or starting an alternative version (inspired by this patch, and with credit given to @EGuesnet for their initial work, in the commit message).

Thanks @jhenderson ! That would work too . @DiggerLin Let us do it -- you can create another patch based on this and give credit to @EGuesnet if he can NOT address the comment in time.

yes, it is no problem for me. thanks .

[EGuesnet]

Hi @jhenderson, @jsji , @DiggerLin,
As I said before, I do not have time to improve this PR before one or two months at least, as our priorities have shifted.
My proposal to @Esme about accepting the patch with minor changes was OK but now there are too much changes requested by reviewers.
So, I accept that you start an improvement / alternative version as long as you credit.

Please tag me when you will open your PR.
For your information, I have opened another PR about writing Big Archive here: https://reviews.llvm.org/D104367 .

[jsji]

In D100651#3060593, @EGuesnet wrote:

Hi @jhenderson, @jsji , @DiggerLin,
As I said before, I do not have time to improve this PR before one or two months at least, as our priorities have shifted.
My proposal to @Esme about accepting the patch with minor changes was OK but now there are too much changes requested by reviewers.
So, I accept that you start an improvement / alternative version as long as you credit.

Please tag me when you will open your PR.
For your information, I have opened another PR about writing Big Archive here: https://reviews.llvm.org/D104367 .

Great! Thank you for your kind and contribution to this! @EGuesnet

@DiggerLin Please credit and tag @EGuesnet in your new patch. Thanks.

llvm/include/llvm/Object/Archive.h

 #include <string>
 #include <vector>
 
 namespace llvm {
 namespace object {
 
 class Archive;
 
-class ArchiveMemberHeader {
+class AbstractArchiveMemberHeader {
 public:
   friend class Archive;
-
-  ArchiveMemberHeader(Archive const *Parent, const char *RawHeaderPtr,
-                      uint64_t Size, Error *Err);
-  // ArchiveMemberHeader() = default;
-
+  // clone() is used to create a new object identical to original.
+  virtual std::unique_ptr<AbstractArchiveMemberHeader> clone() const = 0;
+  virtual ~AbstractArchiveMemberHeader(){};
   /// Get the name without looking up long names.
-  Expected<StringRef> getRawName() const;
+  virtual Expected<StringRef> getRawName() const = 0;
 
   /// Get the name looking up long names.
-  Expected<StringRef> getName(uint64_t Size) const;
+  virtual Expected<StringRef> getName(uint64_t Size) const = 0;
 
-  Expected<uint64_t> getSize() const;
+  virtual Expected<uint64_t> getSize() const = 0;
+
+  // Raw access and helper getters
+  virtual uint64_t getOffset() const = 0;
+  virtual StringRef getRawAccessMode() const = 0;
+  virtual StringRef getRawLastModified() const = 0;
+  virtual StringRef getRawUID() const = 0;
+  virtual StringRef getRawGID() const = 0;
 
+  // Non-Raw getters
   Expected<sys::fs::perms> getAccessMode() const;
   Expected<sys::TimePoint<std::chrono::seconds>> getLastModified() const;
 
-  StringRef getRawLastModified() const {
-    return StringRef(ArMemHdr->LastModified, sizeof(ArMemHdr->LastModified))
-        .rtrim(' ');
-  }
-
   Expected<unsigned> getUID() const;
   Expected<unsigned> getGID() const;
 
   // This returns the size of the private struct ArMemHdrType
-  uint64_t getSizeOf() const { return sizeof(ArMemHdrType); }
+  virtual uint64_t getSizeOf() const = 0;
+  // This returns the size of the private struct ArFixLenHdrType
+  virtual uint64_t getFixSizeOf() const = 0;
+};
+
+class ArchiveMemberHeader : public AbstractArchiveMemberHeader {
+public:
+  ArchiveMemberHeader(Archive const *Parent, const char *RawHeaderPtr,
+                      uint64_t Size, Error *Err);
+  std::unique_ptr<AbstractArchiveMemberHeader> clone() const override {
+    return std::unique_ptr<AbstractArchiveMemberHeader>(
+        new ArchiveMemberHeader(*this));
+  }
+
+  Expected<StringRef> getRawName() const override;
+  Expected<StringRef> getName(uint64_t Size) const override;
+  Expected<uint64_t> getSize() const override;
+
+  uint64_t getOffset() const override;
+  StringRef getRawAccessMode() const override;
+  StringRef getRawLastModified() const override;
+  StringRef getRawUID() const override;
+  StringRef getRawGID() const override;
+
+  uint64_t getSizeOf() const override { return sizeof(ArMemHdrType); }
+  uint64_t getFixSizeOf() const override { return 0; }
 
 private:
   struct ArMemHdrType {
     char Name[16];
     char LastModified[12];
     char UID[6];
     char GID[6];
     char AccessMode[8];
     char Size[10]; ///< Size of data, not including header or padding.
     char Terminator[2];
   };
+  ArMemHdrType const *ArMemHdr;
   Archive const *Parent;
+};
+
+class BigArchiveMemberHeader : public AbstractArchiveMemberHeader {
+public:
+  BigArchiveMemberHeader(Archive const *Parent, const char *RawHeaderPtr,
+                         uint64_t Size, Error *Err);
+  std::unique_ptr<AbstractArchiveMemberHeader> clone() const override {
+    return std::unique_ptr<AbstractArchiveMemberHeader>(
+        new BigArchiveMemberHeader(*this));
+  }
+
+  Expected<StringRef> getRawName() const override;
+  Expected<StringRef> getName(uint64_t Size) const override;
+  Expected<uint64_t> getSize() const override;
+
+  uint64_t getOffset() const override;
+  StringRef getRawAccessMode() const override;
+  StringRef getRawLastModified() const override;
+  StringRef getRawUID() const override;
+  StringRef getRawGID() const override;
+
+  uint64_t getSizeOf() const override { return sizeof(ArMemHdrType); }
+  uint64_t getFixSizeOf() const override { return sizeof(ArFixLenHdrType); }
+
+private:
+  // File Member Header
+  struct ArMemHdrType {
+    char Size[20];       // File member size in decimal
+    char NextOffset[20]; // Next member offset in decimal
+    char PrevOffset[20]; // Previous member offset in decimal
+    char LastModified[12];
+    char UID[12];
+    char GID[12];
+    char AccessMode[12];
+    char NameLen[4]; // File member name length in decimal
+    union {
+      char Name[2]; // Start of member name
+      char Terminator[2];
+    };
+  };
+
+  // Fixed-Length Header (without magic)
+  struct ArFixLenHdrType {
+    char MemOffset[20];       /*Offset to member table */
+    char GlobSymOffset[20];   /*Offset to global symbol table */
+    char GlobSym64Offset[20]; /*Offset global symbol table for 64-bit objects */
+    char FirstArOffset[20];   /*Offset to first archive member */
+    char LastArOffset[20];    /*Offset to last archive member */
+    char FreeOffset[20];      /*Offset to first mem on free list */
+  };
+
   ArMemHdrType const *ArMemHdr;
+  ArFixLenHdrType const *ArFixLenHdr;
+  Archive const *Parent;
 };
 
 class Archive : public Binary {
   virtual void anchor();
 
 public:
   class Child {
     friend Archive;
-    friend ArchiveMemberHeader;
+    friend AbstractArchiveMemberHeader;
 
     const Archive *Parent;
-    ArchiveMemberHeader Header;
+    std::unique_ptr<AbstractArchiveMemberHeader> Header;
     /// Includes header but not padding byte.
     StringRef Data;
     /// Offset from Data to the start of the file.
     uint16_t StartOfFile;
 
     Expected<bool> isThinMember() const;
 
   public:
     Child(const Archive *Parent, const char *Start, Error *Err);
     Child(const Archive *Parent, StringRef Data, uint16_t StartOfFile);
+    Child(const Child &C)
+        : Parent(C.Parent), Data(C.Data), StartOfFile(C.StartOfFile) {
+      if (C.Header)
+        Header = C.Header->clone();
+    }
+    Child(Child &&C) {
+      Parent = std::move(C.Parent);
+      Header = std::move(C.Header);
+      Data = C.Data;
+      StartOfFile = C.StartOfFile;
+    }
+
+    Child &operator=(Child &&C) noexcept {
+      if (&C == this)
+        return *this;
+
+      Parent = std::move(C.Parent);
+      Header = std::move(C.Header);
+      Data = C.Data;
+      StartOfFile = C.StartOfFile;
+
+      return *this;
+    }
+
+    Child &operator=(const Child &C) {
+      if (&C == this)
+        return *this;
+
+      Parent = C.Parent;
+      if (C.Header)
+        Header = C.Header->clone();
+      Data = C.Data;
+      StartOfFile = C.StartOfFile;
+
+      return *this;
+    }
 
     bool operator==(const Child &other) const {
       assert(!Parent || !other.Parent || Parent == other.Parent);
       return Data.begin() == other.Data.begin();
     }
 
     const Archive *getParent() const { return Parent; }
     Expected<Child> getNext() const;
 
     Expected<StringRef> getName() const;
     Expected<std::string> getFullName() const;
-    Expected<StringRef> getRawName() const { return Header.getRawName(); }
+    Expected<StringRef> getRawName() const { return Header->getRawName(); }
 
     Expected<sys::TimePoint<std::chrono::seconds>> getLastModified() const {
-      return Header.getLastModified();
+      return Header->getLastModified();
     }
 
-    StringRef getRawLastModified() const { return Header.getRawLastModified(); }
+    StringRef getRawLastModified() const {
+      return Header->getRawLastModified();
+    }
 
-    Expected<unsigned> getUID() const { return Header.getUID(); }
-    Expected<unsigned> getGID() const { return Header.getGID(); }
+    Expected<unsigned> getUID() const { return Header->getUID(); }
+    Expected<unsigned> getGID() const { return Header->getGID(); }
 
     Expected<sys::fs::perms> getAccessMode() const {
-      return Header.getAccessMode();
+      return Header->getAccessMode();
     }
 
     /// \return the size of the archive member without the header or padding.
     Expected<uint64_t> getSize() const;
     /// \return the size in the archive header for this member.
     Expected<uint64_t> getRawSize() const;
 
     Expected<StringRef> getBuffer() const;
public:
   };
 
   Archive(MemoryBufferRef Source, Error &Err);
   static Expected<std::unique_ptr<Archive>> create(MemoryBufferRef Source);
 
   /// Size field is 10 decimal digits long
   static const uint64_t MaxMemberSize = 9999999999;
 
-  enum Kind { K_GNU, K_GNU64, K_BSD, K_DARWIN, K_DARWIN64, K_COFF };
+  enum Kind { K_GNU, K_GNU64, K_BSD, K_DARWIN, K_DARWIN64, K_COFF, K_AIXBIG };
 
   Kind kind() const { return (Kind)Format; }
   bool isThin() const { return IsThin; }
 
   child_iterator child_begin(Error &Err, bool SkipInternal = true) const;
   child_iterator child_end() const;
   iterator_range<child_iterator> children(Error &Err,
                                           bool SkipInternal = true) const {
     return make_range(child_begin(Err, SkipInternal), child_end());
   }
 
   symbol_iterator symbol_begin() const;
   symbol_iterator symbol_end() const;
   iterator_range<symbol_iterator> symbols() const {
public:
   StringRef getSymbolTable() const { return SymbolTable; }
   StringRef getStringTable() const { return StringTable; }
   uint32_t getNumberOfSymbols() const;
 
   std::vector<std::unique_ptr<MemoryBuffer>> takeThinBuffers() {
     return std::move(ThinBuffers);
   }
 
+  // On AIX Big Archive, total length is needed
+  // because end of file is member table and global symbol table.
+  uint64_t Length = 0;
+  // On AIX Big Archive, all offset are global offset,
+  // so we need to memorize position.
+  static uint64_t CurrentLocation;
+
 private:
   StringRef SymbolTable;
   StringRef StringTable;
 
   StringRef FirstRegularData;
   uint16_t FirstRegularStartOfFile = -1;
   void setFirstRegular(const Child &C);
 
   unsigned Format : 3;
   unsigned IsThin : 1;
   mutable std::vector<std::unique_ptr<MemoryBuffer>> ThinBuffers;
 };
 
 } // end namespace object
 } // end namespace llvm
 
 #endif // LLVM_OBJECT_ARCHIVE_H

llvm/lib/Object/Archive.cpp

 #include <system_error>
 
 using namespace llvm;
 using namespace object;
 using namespace llvm::support::endian;
 
 const char Magic[] = "!<arch>\n";
 const char ThinMagic[] = "!<thin>\n";
+const char BigMagic[] = "<bigaf>\n";
+
+// All magic are 8 caractere long
+#define MAGIC_LEN (uint)8
+
+uint64_t Archive::CurrentLocation = 0;
 
 void Archive::anchor() {}
 
 static Error malformedError(Twine Msg) {
   std::string StringMsg = "truncated or malformed archive (" + Msg.str() + ")";
   return make_error<GenericBinaryError>(std::move(StringMsg),
                                         object_error::parse_failed);
 }
 
 ArchiveMemberHeader::ArchiveMemberHeader(const Archive *Parent,
                                          const char *RawHeaderPtr,
                                          uint64_t Size, Error *Err)
-    : Parent(Parent),
-      ArMemHdr(reinterpret_cast<const ArMemHdrType *>(RawHeaderPtr)) {
+    : Parent(Parent) {
   if (RawHeaderPtr == nullptr)
     return;
   ErrorAsOutParameter ErrAsOutParam(Err);
 
+  ArMemHdr = reinterpret_cast<const ArMemHdrType *>(RawHeaderPtr);
+
   if (Size < sizeof(ArMemHdrType)) {
     if (Err) {
       std::string Msg("remaining size of archive too small for next archive "
                       "member header ");
       Expected<StringRef> NameOrErr = getName(Size);
       if (!NameOrErr) {
         consumeError(NameOrErr.takeError());
         uint64_t Offset = RawHeaderPtr - Parent->getData().data();
if (Err) {
         *Err = malformedError(Msg + "at offset " + Twine(Offset));
       } else
         *Err = malformedError(Msg + "for " + NameOrErr.get());
     }
     return;
   }
 }
 
+BigArchiveMemberHeader::BigArchiveMemberHeader(const Archive *Parent,
+                                               const char *RawHeaderPtr,
+                                               uint64_t Size, Error *Err)
+    : Parent(Parent) {
+  if (RawHeaderPtr == nullptr)
+    return;
+  ErrorAsOutParameter ErrAsOutParam(Err);
+
+  // Begin of the archive: check if it is Fixed-Length Header.
+  if (StringRef(RawHeaderPtr - MAGIC_LEN, MAGIC_LEN).startswith(BigMagic)) {
+    // Fixed-Length Header first
+    ArFixLenHdr = reinterpret_cast<const ArFixLenHdrType *>(RawHeaderPtr);
+    // File-Member Header
+    const char *RawMemberHeaderPtr = RawHeaderPtr + sizeof(ArFixLenHdrType);
+    ArMemHdr = reinterpret_cast<const ArMemHdrType *>(RawMemberHeaderPtr);
+  } else {
+    // Only File-Member Header
+    ArMemHdr = reinterpret_cast<const ArMemHdrType *>(RawHeaderPtr);
+  }
+
+  if (Size < sizeof(ArMemHdrType)) {
+    if (Err) {
+      std::string Msg("remaining size of archive too small for next archive "
+                      "member header ");
+      Expected<StringRef> NameOrErr = getName(Size);
+      if (!NameOrErr) {
+        consumeError(NameOrErr.takeError());
+        uint64_t Offset = RawHeaderPtr - Parent->getData().data();
+        *Err = malformedError(Msg + "at offset " + Twine(Offset));
+      } else
+        *Err = malformedError(Msg + "for " + NameOrErr.get());
+    }
+    return;
+  }
+
+  // Terminator is cosmetic only for AIX Big Archive
+}
+
 // This gets the raw name from the ArMemHdr->Name field and checks that it is
 // valid for the kind of archive.  If it is not valid it returns an Error.
 Expected<StringRef> ArchiveMemberHeader::getRawName() const {
   char EndCond;
   auto Kind = Parent->kind();
   if (Kind == Archive::K_BSD || Kind == Archive::K_DARWIN64) {
     if (ArMemHdr->Name[0] == ' ') {
       uint64_t Offset =
Expected<StringRef> ArchiveMemberHeader::getRawName() const {
   if (end == StringRef::npos)
     end = sizeof(ArMemHdr->Name);
   assert(end <= sizeof(ArMemHdr->Name) && end > 0);
   // Don't include the EndCond if there is one.
   return StringRef(ArMemHdr->Name, end);
 }
 
 // This gets the name looking up long names. Size is the size of the archive
+Expected<StringRef> BigArchiveMemberHeader::getRawName() const {
+  // Name is outside ArMemHdr, and there is no end caracter
+  // name lenght is in NameLen field
+  // The two first char of name are already in ArMemHdrType
+  // but unused terminator '`\n' is after the name.
+  StringRef::size_type end = strtol(ArMemHdr->NameLen, NULL, 10);
+  return StringRef(ArMemHdr->Name, end);
+}
+
 // member including the header, so the size of any name following the header
 // is checked to make sure it does not overflow.
 Expected<StringRef> ArchiveMemberHeader::getName(uint64_t Size) const {
 
   // This can be called from the ArchiveMemberHeader constructor when the
   // archive header is truncated to produce an error message with the name.
   // Make sure the name field is not truncated.
   if (Size < offsetof(ArMemHdrType, Name) + sizeof(ArMemHdr->Name)) {
Expected<StringRef> ArchiveMemberHeader::getName(uint64_t Size) const {
   // It is not a long name so trim the blanks at the end of the name.
   if (Name[Name.size() - 1] != '/')
     return Name.rtrim(' ');
 
   // It's a simple name.
   return Name.drop_back(1);
 }
 
+Expected<StringRef> BigArchiveMemberHeader::getName(uint64_t Size) const {
+  // The raw name itself can be invalid.
+  Expected<StringRef> NameOrErr = getRawName();
+  if (!NameOrErr)
+    return NameOrErr.takeError();
+  StringRef Name = NameOrErr.get();
+
+  // Trim the blanks at the end of the name.
+  return Name.rtrim(' ');
+}
+
 Expected<uint64_t> ArchiveMemberHeader::getSize() const {
-  uint64_t Ret;
+  uint64_t Size;
   if (StringRef(ArMemHdr->Size, sizeof(ArMemHdr->Size))
           .rtrim(" ")
-          .getAsInteger(10, Ret)) {
+          .getAsInteger(10, Size)) {
     std::string Buf;
     raw_string_ostream OS(Buf);
     OS.write_escaped(
         StringRef(ArMemHdr->Size, sizeof(ArMemHdr->Size)).rtrim(" "));
     OS.flush();
     uint64_t Offset =
         reinterpret_cast<const char *>(ArMemHdr) - Parent->getData().data();
     return malformedError("characters in size field in archive header are not "
                           "all decimal numbers: '" +
                           Buf +
                           "' for archive "
                           "member header at offset " +
                           Twine(Offset));
   }
-  return Ret;
+  return Size;
 }
 
-Expected<sys::fs::perms> ArchiveMemberHeader::getAccessMode() const {
-  unsigned Ret;
-  if (StringRef(ArMemHdr->AccessMode, sizeof(ArMemHdr->AccessMode))
-          .rtrim(' ')
-          .getAsInteger(8, Ret)) {
+Expected<uint64_t> BigArchiveMemberHeader::getSize() const {
+  uint64_t Size;
+  uint64_t NameLen;
+  if (StringRef(ArMemHdr->Size, sizeof(ArMemHdr->Size))
+          .rtrim(" ")
+          .getAsInteger(10, Size)) {
     std::string Buf;
     raw_string_ostream OS(Buf);
     OS.write_escaped(
-        StringRef(ArMemHdr->AccessMode, sizeof(ArMemHdr->AccessMode))
-            .rtrim(" "));
+        StringRef(ArMemHdr->Size, sizeof(ArMemHdr->Size)).rtrim(" "));
     OS.flush();
     uint64_t Offset =
         reinterpret_cast<const char *>(ArMemHdr) - Parent->getData().data();
+    return malformedError("characters in size field in archive header are not "
+                          "all decimal numbers: '" +
+                          Buf +
+                          "' for archive "
+                          "member header at offset " +
+                          Twine(Offset));
+  }
+  if (StringRef(ArMemHdr->NameLen, sizeof(ArMemHdr->NameLen))
+          .rtrim(" ")
+          .getAsInteger(10, NameLen)) {
+    std::string Buf;
+    raw_string_ostream OS(Buf);
+    OS.write_escaped(
+        StringRef(ArMemHdr->NameLen, sizeof(ArMemHdr->NameLen)).rtrim(" "));
+    OS.flush();
+    uint64_t Offset =
+        reinterpret_cast<const char *>(ArMemHdr) - Parent->getData().data();
+    return malformedError(
+        "characters in name length field in archive header are not "
+        "all decimal numbers: '" +
+        Buf +
+        "' for archive "
+        "member header at offset " +
+        Twine(Offset));
+  }
+  // First read: size is header + object size + name round to be even +
+  // Fixed-Length Header + magic header is added automatically
+  if (Parent->CurrentLocation == MAGIC_LEN)
+    return Size + NameLen + NameLen % 2 + sizeof(ArFixLenHdrType);
+  // Next read: size is header + object size + name round to be even
+  else
+    return Size + NameLen + NameLen % 2;
+}
+
+uint64_t ArchiveMemberHeader::getOffset() const {
+  uint64_t Offset =
+      reinterpret_cast<const char *>(ArMemHdr) - Parent->getData().data();
+  return Offset;
+}
+
+uint64_t BigArchiveMemberHeader::getOffset() const {
+  uint64_t Offset =
+      reinterpret_cast<const char *>(ArMemHdr) - Parent->getData().data();
+  return Offset;
+}
+
+// This gets the raw access mode from the ArMemHdr->AccessMode field.
+StringRef ArchiveMemberHeader::getRawAccessMode() const {
+  return StringRef(ArMemHdr->AccessMode, sizeof(ArMemHdr->AccessMode))
+      .rtrim(' ');
+}
+
+StringRef BigArchiveMemberHeader::getRawAccessMode() const {
+  return StringRef(ArMemHdr->AccessMode, sizeof(ArMemHdr->AccessMode))
+      .rtrim(' ');
+}
+
+Expected<sys::fs::perms> AbstractArchiveMemberHeader::getAccessMode() const {
+  unsigned Ret;
+  if (getRawAccessMode().getAsInteger(8, Ret)) {
+    std::string Buf;
+    raw_string_ostream OS(Buf);
+    OS.write_escaped(getRawAccessMode());
+    OS.flush();
+    uint64_t Offset = getOffset();
     return malformedError("characters in AccessMode field in archive header "
                           "are not all decimal numbers: '" +
                           Buf + "' for the archive member header at offset " +
                           Twine(Offset));
   }
   return static_cast<sys::fs::perms>(Ret);
 }
 
+// This gets ArMemHdr->LastModified field.
+StringRef ArchiveMemberHeader::getRawLastModified() const {
+  return StringRef(ArMemHdr->LastModified, sizeof(ArMemHdr->LastModified))
+      .rtrim(' ');
+}
+
+StringRef BigArchiveMemberHeader::getRawLastModified() const {
+  return StringRef(ArMemHdr->LastModified, sizeof(ArMemHdr->LastModified))
+      .rtrim(' ');
+}
+
 Expected<sys::TimePoint<std::chrono::seconds>>
-ArchiveMemberHeader::getLastModified() const {
+AbstractArchiveMemberHeader::getLastModified() const {
   unsigned Seconds;
-  if (StringRef(ArMemHdr->LastModified, sizeof(ArMemHdr->LastModified))
-          .rtrim(' ')
-          .getAsInteger(10, Seconds)) {
+  if (getRawLastModified().getAsInteger(10, Seconds)) {
     std::string Buf;
     raw_string_ostream OS(Buf);
-    OS.write_escaped(
-        StringRef(ArMemHdr->LastModified, sizeof(ArMemHdr->LastModified))
-            .rtrim(" "));
+    OS.write_escaped(StringRef(getRawLastModified()));
     OS.flush();
-    uint64_t Offset =
-        reinterpret_cast<const char *>(ArMemHdr) - Parent->getData().data();
+    uint64_t Offset = getOffset();
     return malformedError("characters in LastModified field in archive header "
                           "are not all decimal numbers: '" +
                           Buf + "' for the archive member header at offset " +
                           Twine(Offset));
   }
 
   return sys::toTimePoint(Seconds);
 }
 
-Expected<unsigned> ArchiveMemberHeader::getUID() const {
+// This gets the raw UID from the ArMemHdr->UID field.
+StringRef ArchiveMemberHeader::getRawUID() const {
+  return StringRef(ArMemHdr->UID, sizeof(ArMemHdr->UID)).rtrim(' ');
+}
+
+StringRef BigArchiveMemberHeader::getRawUID() const {
+  return StringRef(ArMemHdr->UID, sizeof(ArMemHdr->UID)).rtrim(' ');
+}
+
+Expected<unsigned> AbstractArchiveMemberHeader::getUID() const {
   unsigned Ret;
-  StringRef User = StringRef(ArMemHdr->UID, sizeof(ArMemHdr->UID)).rtrim(' ');
+  StringRef User = getRawUID();
   if (User.empty())
     return 0;
   if (User.getAsInteger(10, Ret)) {
     std::string Buf;
     raw_string_ostream OS(Buf);
     OS.write_escaped(User);
     OS.flush();
-    uint64_t Offset =
-        reinterpret_cast<const char *>(ArMemHdr) - Parent->getData().data();
+    uint64_t Offset = getOffset();
     return malformedError("characters in UID field in archive header "
                           "are not all decimal numbers: '" +
                           Buf + "' for the archive member header at offset " +
                           Twine(Offset));
   }
   return Ret;
 }
 
-Expected<unsigned> ArchiveMemberHeader::getGID() const {
+// This gets the raw GID from the ArMemHdr->GID field.
+StringRef ArchiveMemberHeader::getRawGID() const {
+  return StringRef(ArMemHdr->GID, sizeof(ArMemHdr->GID)).rtrim(' ');
+}
+
+StringRef BigArchiveMemberHeader::getRawGID() const {
+  return StringRef(ArMemHdr->GID, sizeof(ArMemHdr->GID)).rtrim(' ');
+}
+
+Expected<unsigned> AbstractArchiveMemberHeader::getGID() const {
   unsigned Ret;
-  StringRef Group = StringRef(ArMemHdr->GID, sizeof(ArMemHdr->GID)).rtrim(' ');
+  StringRef Group = getRawGID();
   if (Group.empty())
     return 0;
   if (Group.getAsInteger(10, Ret)) {
     std::string Buf;
     raw_string_ostream OS(Buf);
     OS.write_escaped(Group);
     OS.flush();
-    uint64_t Offset =
-        reinterpret_cast<const char *>(ArMemHdr) - Parent->getData().data();
+    uint64_t Offset = getOffset();
     return malformedError("characters in GID field in archive header "
                           "are not all decimal numbers: '" +
                           Buf + "' for the archive member header at offset " +
                           Twine(Offset));
   }
   return Ret;
 }
 
 Archive::Child::Child(const Archive *Parent, StringRef Data,
                       uint16_t StartOfFile)
-    : Parent(Parent), Header(Parent, Data.data(), Data.size(), nullptr),
-      Data(Data), StartOfFile(StartOfFile) {}
+    : Parent(Parent), Data(Data), StartOfFile(StartOfFile) {
+  // Create the right concrete archive member as a function of Kind.
+  if (Parent->kind() != K_AIXBIG) {
+    Header = std::make_unique<ArchiveMemberHeader>(
+        ArchiveMemberHeader(Parent, Data.data(), Data.size(), nullptr));
+  } else {
+    Header = std::make_unique<BigArchiveMemberHeader>(
+        BigArchiveMemberHeader(Parent, Data.data(), Data.size(), nullptr));
+  }
+}
 
 Archive::Child::Child(const Archive *Parent, const char *Start, Error *Err)
-    : Parent(Parent),
-      Header(Parent, Start,
-             Parent
-                 ? Parent->getData().size() - (Start - Parent->getData().data())
-                 : 0,
-             Err) {
-  if (!Start)
+    : Parent(Parent) {
+  if (!Start) {
+    Header = nullptr;
     return;
+  }
+
+  // Create the right concrete archive member as a function of Kind.
+  if (Parent->kind() != K_AIXBIG) {
+    Header = std::make_unique<ArchiveMemberHeader>(ArchiveMemberHeader(
+        Parent, Start,
+        Parent ? Parent->getData().size() - (Start - Parent->getData().data())
+               : 0,
+        Err));
+  } else {
+    Archive::CurrentLocation = Start - Parent->getData().data();
+    Header = std::make_unique<BigArchiveMemberHeader>(BigArchiveMemberHeader(
+        Parent, Start,
+        Parent ? Parent->getData().size() - (Start - Parent->getData().data())
+               : 0,
+        Err));
+  }
 
   // If we are pointed to real data, Start is not a nullptr, then there must be
   // a non-null Err pointer available to report malformed data on.  Only in
   // the case sentinel value is being constructed is Err is permitted to be a
   // nullptr.
   assert(Err && "Err can't be nullptr if Start is not a nullptr");
 
   ErrorAsOutParameter ErrAsOutParam(Err);
 
   // If there was an error in the construction of the Header
   // then just return with the error now set.
   if (*Err)
     return;
 
-  uint64_t Size = Header.getSizeOf();
+  uint64_t Size = Header->getSizeOf();
   Data = StringRef(Start, Size);
   Expected<bool> isThinOrErr = isThinMember();
   if (!isThinOrErr) {
     *Err = isThinOrErr.takeError();
     return;
   }
   bool isThin = isThinOrErr.get();
   if (!isThin) {
     Expected<uint64_t> MemberSize = getRawSize();
     if (!MemberSize) {
       *Err = MemberSize.takeError();
       return;
     }
     Size += MemberSize.get();
     Data = StringRef(Start, Size);
   }
 
   // Setup StartOfFile and PaddingBytes.
-  StartOfFile = Header.getSizeOf();
+  StartOfFile = Header->getSizeOf();
   // Don't include attached name.
   Expected<StringRef> NameOrErr = getRawName();
   if (!NameOrErr) {
     *Err = NameOrErr.takeError();
     return;
   }
   StringRef Name = NameOrErr.get();
+
+  if (Parent->kind() == Archive::K_AIXBIG &&
+      Parent->CurrentLocation <= Header->getFixSizeOf() + MAGIC_LEN) {
+    // Add name to found the real start
+    // Add also Fixed-Length Header in the first read.
+    StartOfFile += Name.size() + Name.size() % 2;
+    StartOfFile += Header->getFixSizeOf();
+  } else if (Parent->kind() == Archive::K_AIXBIG) {
+    // Add name to found the real start
+    StartOfFile += Name.size() + Name.size() % 2;
+  }
+
   if (Name.startswith("#1/")) {
     uint64_t NameSize;
     if (Name.substr(3).rtrim(' ').getAsInteger(10, NameSize)) {
       std::string Buf;
       raw_string_ostream OS(Buf);
       OS.write_escaped(Name.substr(3).rtrim(' '));
       OS.flush();
       uint64_t Offset = Start - Parent->getData().data();
       *Err = malformedError("long name length characters after the #1/ are "
                             "not all decimal numbers: '" +
                             Buf + "' for archive member header at offset " +
                             Twine(Offset));
       return;
     }
     StartOfFile += NameSize;
   }
 }
 
 Expected<uint64_t> Archive::Child::getSize() const {
   if (Parent->IsThin)
-    return Header.getSize();
+    return Header->getSize();
   return Data.size() - StartOfFile;
 }
 
 Expected<uint64_t> Archive::Child::getRawSize() const {
-  return Header.getSize();
+  return Header->getSize();
 }
 
 Expected<bool> Archive::Child::isThinMember() const {
-  Expected<StringRef> NameOrErr = Header.getRawName();
+  Expected<StringRef> NameOrErr = Header->getRawName();
   if (!NameOrErr)
     return NameOrErr.takeError();
   StringRef Name = NameOrErr.get();
   return Parent->IsThin && Name != "/" && Name != "//";
 }
 
 Expected<std::string> Archive::Child::getFullName() const {
   Expected<bool> isThin = isThinMember();
Expected<StringRef> Archive::Child::getBuffer() const {
   const std::string &FullName = *FullNameOrErr;
   ErrorOr<std::unique_ptr<MemoryBuffer>> Buf = MemoryBuffer::getFile(FullName);
   if (std::error_code EC = Buf.getError())
     return errorCodeToError(EC);
   Parent->ThinBuffers.push_back(std::move(*Buf));
   return Parent->ThinBuffers.back()->getBuffer();
 }
 
 Expected<Archive::Child> Archive::Child::getNext() const {
   size_t SpaceToSkip = Data.size();
   // If it's odd, add 1 to make it even.
   if (SpaceToSkip & 1)
     ++SpaceToSkip;
 
   const char *NextLoc = Data.data() + SpaceToSkip;
+  // Update current; will be updated in Child constructor
+  const uint64_t CurrentLocation = Parent->CurrentLocation + SpaceToSkip;
 
   // Check to see if this is at the end of the archive.
-  if (NextLoc == Parent->Data.getBufferEnd())
+  // With BigArchive, stop on Member table.
+  if (NextLoc == Parent->Data.getBufferEnd() ||
+      (Parent->kind() == K_AIXBIG && CurrentLocation == Parent->Length)) {
     return Child(nullptr, nullptr, nullptr);
+  }
 
   // Check to see if this is past the end of the archive.
-  if (NextLoc > Parent->Data.getBufferEnd()) {
+  // With BigArchive, stop on Member table.
+  if ((NextLoc > Parent->Data.getBufferEnd()) ||
+      (Parent->kind() == K_AIXBIG && CurrentLocation > Parent->Length)) {
     std::string Msg("offset to next archive member past the end of the archive "
                     "after member ");
     Expected<StringRef> NameOrErr = getName();
     if (!NameOrErr) {
       consumeError(NameOrErr.takeError());
       uint64_t Offset = Data.data() - Parent->getData().data();
       return malformedError(Msg + "at offset " + Twine(Offset));
     } else
uint64_t Archive::Child::getChildOffset() const {
   return offset;
 }
 
 Expected<StringRef> Archive::Child::getName() const {
   Expected<uint64_t> RawSizeOrErr = getRawSize();
   if (!RawSizeOrErr)
     return RawSizeOrErr.takeError();
   uint64_t RawSize = RawSizeOrErr.get();
-  Expected<StringRef> NameOrErr = Header.getName(Header.getSizeOf() + RawSize);
+  Expected<StringRef> NameOrErr =
+      Header->getName(Header->getSizeOf() + RawSize);
   if (!NameOrErr)
     return NameOrErr.takeError();
   StringRef Name = NameOrErr.get();
   return Name;
 }
 
 Expected<MemoryBufferRef> Archive::Child::getMemoryBufferRef() const {
   Expected<StringRef> NameOrErr = getName();
   if (!NameOrErr)
Archive::Archive(MemoryBufferRef Source, Error &Err)
     : Binary(Binary::ID_Archive, Source) {
   ErrorAsOutParameter ErrAsOutParam(&Err);
   StringRef Buffer = Data.getBuffer();
   // Check for sufficient magic.
   if (Buffer.startswith(ThinMagic)) {
     IsThin = true;
   } else if (Buffer.startswith(Magic)) {
     IsThin = false;
+  } else if (Buffer.startswith(BigMagic)) {
+    IsThin = false;
   } else {
     Err = make_error<GenericBinaryError>("file too small to be an archive",
                                          object_error::invalid_file_type);
     return;
   }
 
   // Make sure Format is initialized before any call to
-  // ArchiveMemberHeader::getName() is made.  This could be a valid empty
-  // archive which is the same in all formats.  So claiming it to be gnu to is
-  // fine if not totally correct before we look for a string table or table of
-  // contents.
+  // ArchiveMemberHeader::getName() is made. Read Magic to deal with AIX Big
+  // Archive. Else, this could be a valid empty archive which is the same in all
+  // formats. So claiming it to be gnu to is fine if not totally correct before
+  // we look for a string table or table of contents.
+  if (Buffer.startswith(BigMagic)) {
+    Format = K_AIXBIG;
+  } else
     Format = K_GNU;
 
   // Get the special members.
   child_iterator I = child_begin(Err, false);
   if (Err)
     return;
   child_iterator E = child_end();
 
   // See if this is a valid empty archive and if so return.
Archive::Archive(MemoryBufferRef Source, Error &Err)
 
   Expected<StringRef> NameOrErr = C->getRawName();
   if (!NameOrErr) {
     Err = NameOrErr.takeError();
     return;
   }
   StringRef Name = NameOrErr.get();
 
+  // AIX Big Archive format
+  //   Identified purely by magic bytes and uses a unique format.
+  if (Buffer.startswith(BigMagic)) {
+    Format = K_AIXBIG;
+
+    // Length of all object file and their headers corresponds to "offset to
+    // member table", located at 8->27 in fixed length header.
+    Buffer.substr(8, 20).rtrim(" ").getAsInteger(10, Length);
+    setFirstRegular(*C);
+    Err = Error::success();
+    return;
+  }
+
   // Below is the pattern that is used to figure out the archive format
   // GNU archive format
   //  First member : / (may exist, if it exists, points to the symbol table )
   //  Second member : // (may exist, if it exists, points to the string table)
   //  Note : The string table is used if the filename exceeds 15 characters
   // BSD archive format
   //  First member : __.SYMDEF or "__.SYMDEF SORTED" (the symbol table)
   //  There is no string table, if the filename exceeds 15 characters or has a
Archive::child_iterator Archive::child_begin(Error &Err,
   if (Err)
     return child_end();
   return child_iterator::itr(C, Err);
 }
 
 Archive::child_iterator Archive::child_end() const {
   return child_iterator::end(Child(nullptr, nullptr, nullptr));
 }
 
 StringRef Archive::Symbol::getName() const {
   return Parent->getSymbolTable().begin() + StringIndex;
 }
 
 Expected<Archive::Child> Archive::Symbol::getMember() const {
   const char *Buf = Parent->getSymbolTable().begin();
   const char *Offsets = Buf;
   if (Parent->kind() == K_GNU64 || Parent->kind() == K_DARWIN64)
     Offsets += sizeof(uint64_t);
   else
     Offsets += sizeof(uint32_t);
   uint64_t Offset = 0;
   if (Parent->kind() == K_GNU) {
     Offset = read32be(Offsets + SymbolIndex * 4);
   } else if (Parent->kind() == K_GNU64) {
     Offset = read64be(Offsets + SymbolIndex * 8);
   } else if (Parent->kind() == K_BSD) {
     // The SymbolIndex is an index into the ranlib structs that start at
     // Offsets (the first uint32_t is the number of bytes of the ranlib
     // structs).  The ranlib structs are a pair of uint32_t's the first
     // being a string table offset and the second being the offset into

llvm/lib/Object/ArchiveWriter.cpp

static bool isBSDLike(object::Archive::Kind Kind) {
   switch (Kind) {
   case object::Archive::K_GNU:
   case object::Archive::K_GNU64:
     return false;
   case object::Archive::K_BSD:
   case object::Archive::K_DARWIN:
   case object::Archive::K_DARWIN64:
     return true;
+  case object::Archive::K_AIXBIG:
+    report_fatal_error("Not handled yet");
   case object::Archive::K_COFF:
     break;
   }
   llvm_unreachable("not supported for writting");
 }
 
 template <class T>
 static void print(raw_ostream &Out, object::Archive::Kind Kind, T Val) {
static bool is64BitKind(object::Archive::Kind Kind) {
   case object::Archive::K_GNU:
   case object::Archive::K_BSD:
   case object::Archive::K_DARWIN:
   case object::Archive::K_COFF:
     return false;
   case object::Archive::K_DARWIN64:
   case object::Archive::K_GNU64:
     return true;
+  case object::Archive::K_AIXBIG:
+    llvm_unreachable("Not handled yet");
   }
   llvm_unreachable("not supported for writting");
 }
 
 static void
 printMemberHeader(raw_ostream &Out, uint64_t Pos, raw_ostream &StringTable,
                   StringMap<uint64_t> &MemberNames, object::Archive::Kind Kind,
                   bool Thin, const NewArchiveMember &M,

llvm/test/Object/archive-big-extract.test

+## Test extracting an archive created by AIX ar (Big Archive).
+RUN: llvm-ar p %p/Inputs/bigFormat.a evenlen | FileCheck %s --check-prefix=AIXBIG --implicit-check-not={{.}}
+AIXBIG: content_of_evenlen

llvm/test/Object/archive-big-read.test

+## Test reading an archive created by AIX ar (Big Archive).
+RUN: env TZ=GMT llvm-ar tv %p/Inputs/bigFormat.a | FileCheck %s --strict-whitespace --check-prefix=AIXBIG --implicit-check-not={{.}}
+AIXBIG:      rw-r--r-- 0/0     19 Jan  1 00:00 1970 evenlen
+AIXBIG-NEXT: rw-r--r-- 0/0      7 Jan  1 00:00 1970 oddlen
+AIXBIG-NEXT: rw-r--r-- 0/0    258 Jan  1 00:00 1970 empty.o