[llvm-objdump] Create name for fake sections

[namhyung]

Summary

It doesn't have a section header string table so add a vector to have
the strings and create name based on the program header type and the
index.

[jhenderson]

Two thoughts:

1. Have you looked into using StringTableBuilder at all? I think you should be able to use that instead of hand-rolling your own code to achieve what looks to be the same thing.
2. I'm not convinced by the naming scheme for the fake section names. Are you following any existing precedent for it? If not, I think we need it to be more obviously different to regular section names, as it doesn't seem unreasonable for people to have custom section names like "load0" for their sections. My immediate thought was something like PT_LOAD#0 for e.g. the PT_LOAD at index 0.

@MaskRay, any thoughts on the naming scheme?

[namhyung]

1. Have you looked into using StringTableBuilder at all?

Nope, will take a look!

2. Are you following any existing precedent for it?

Yes, it follows binutils objdump.

[MaskRay]

Does GNU objdump use load0?

% objdump -h llvm/test/Object/Inputs/no-sections.elf-x86-64 

llvm/test/Object/Inputs/no-sections.elf-x86-64:     file format elf64-x86-64

Sections:
Idx Name          Size      VMA               LMA               File off  Algn

I think PT_LOAD#0 looks good to me, but I don't object to matching GNU objdump for the commonly used PT_LOAD.
There is a small issue that we have to invent names for PT_* values, for values that GNU objdump doesn't hard code.
This is probably an area whether 100% compatibility doesn't matter.

llvm/include/llvm/Object/ELF.h

186

Use SmallString<0> FakeSectionStrings.

Prefer SmallString because std::string has a small string optimization which isn't useful for this case.

[namhyung]

Not sure what happened with the input data. But at least it worked with my test case and linux perf kcore image.

$ bin/yaml2obj ../llvm/test/tools/llvm-objdump/X86/disassemble-no-section.test -o nosec.elf

$ objdump -h nosec.elf

nosec.elf:     file format elf64-x86-64

Sections:
Idx Name          Size      VMA               LMA               File off  Algn
  0 load0         0000000a  ffffffff00000000  ffffffff00000000  00001000  2**0
                  CONTENTS, ALLOC, LOAD, READONLY, CODE

$ objdump -d nosec.elf

nosec.elf:     file format elf64-x86-64


Disassembly of section load0:

ffffffff00000000 <load0>:
ffffffff00000000:	55                   	push   %rbp
ffffffff00000001:	48 89 e5             	mov    %rsp,%rbp
ffffffff00000004:	0f 1f 40 00          	nopl   0x0(%rax)
ffffffff00000008:	5d                   	pop    %rbp
ffffffff00000009:	c3                   	ret    

[namhyung]

llvm/include/llvm/Object/ELF.h

186

But I think it'd be better to use StringTableBuilder as @jhenderson said.

[jhenderson]

In D131290#3710905, @MaskRay wrote:

Does GNU objdump use load0?

% objdump -h llvm/test/Object/Inputs/no-sections.elf-x86-64 

llvm/test/Object/Inputs/no-sections.elf-x86-64:     file format elf64-x86-64

Sections:
Idx Name          Size      VMA               LMA               File off  Algn

I think PT_LOAD#0 looks good to me, but I don't object to matching GNU objdump for the commonly used PT_LOAD.
There is a small issue that we have to invent names for PT_* values, for values that GNU objdump doesn't hard code.
This is probably an area whether 100% compatibility doesn't matter.

I'm not sure we need to worry about inventing names for other PT_* values at this time, because fake sections are only created for executable PT_LOAD, if I'm not mistaken. That being said, I agree that it's unlikely it matters whether this is 100% compatible.

[namhyung]

Updates

• use StringTableBuilder and SmallString
• fix bug in the p_type check

[namhyung]

llvm/include/llvm/Object/ELF.h

778–780

I've found a bug in this code - it should check equality.

[jhenderson]

I was just experimenting with an admittedly old version of GNU objdump on my machine and that just tells me that there are no sections in the file and doesn't do any disassembly. It's possible it's just because it's that old version (I don't have a more recent version available currently). If this is available in GNU objdump, could you post the version it was added to, please?

That said, I think there is benefit from this regardless of the parity with GNU.

llvm/include/llvm/Object/ELF.h

778

This probably wants to be a size_t since it's an index.

779

Not sure you need the llvm:: prefixes?

788

llvm/test/Object/objdump-no-sectionheaders.test

0–59

I'd strongly support replacing the pre-canned ELF with a YAML input in this test. At the moment, I have no idea if the behaviour is correct without digging out a binary inspection tool to look at the ELF object.

Some key characteristics I think that need testing:

1. That non-PT_LOAD segments are ignored (but still count towards the index), even if they have PF_X.
2. That non-PF_X segments are ignored (but still count towards the index).
3. That multiple PT_LOAD wiht PF_X segments can be displayed.

To achieve these points, I think it would be sufficient to have an ELF with the following program headers:

1. PT_PHDR (or other PT_* type) with PF_X
2. PT_LOAD with PF_X
3. PT_LOAD without PF_X
4. PT_LOAD with PF_X

[namhyung]

If this is available in GNU objdump, could you post the version it was added to, please?

I don't know when it's added, it's hard to find.. But my version is 2.38.

namhyung@youngsil:llvm-project$ objdump --version
GNU objdump (GNU Binutils for Debian) 2.38
Copyright (C) 2022 Free Software Foundation, Inc.
This program is free software; you may redistribute it under the terms of
the GNU General Public License version 3 or (at your option) any later version.
This program has absolutely no warranty.

llvm/include/llvm/Object/ELF.h

778

Ok.

779

It seems not. Will remove.

llvm/test/Object/objdump-no-sectionheaders.test

0–59

Sounds good!

[MaskRay]

llvm/include/llvm/Object/ELF.h

186

I think StringTableBuilder just adds complexity/overhead in this case. StringTableBuilder cannot deduplicate any string here (and the code does not need to optimize for size).

653

// There is no section name string table. Return FakeSectionStrings which is non-empty if we have created fake sections.

654

Just return FakeSectionStrings; and delete return ""

778–780

The convention does not place parens around !=

[jhenderson]

llvm/include/llvm/Object/ELF.h

186

I disagree that it adds complexity. It seems more intuitive to do StrTab.add(someString) than std::copy(someString.begin(), someString.end(), std::back_inserter(StrTab)); followed by a null terminator addition. I do agree that the optimizations are unnecessary. Maybe there's room for a "simple" string builder class that does none of the deduplication/optimization?

[namhyung]

llvm/include/llvm/Object/ELF.h

186

I agree with @jhenderson. The API in the StringTableBuilder is simpler and more intuitive to me. Also finalizeInOrder() is even not trying to optimize it.

778–780

Yeah, but I thought it'd good to be symmetric to the RHS.

[namhyung]

Updated.

• remove unnecessary llvm:: prefix
• convert disassemble-no-section.test to use yaml2obj

Note that the test fails due to a mismatch on the first section name.
It should be PT_LOAD#1 but it shows PT_LOAD#3. I don't know why. :(

[jhenderson]

llvm/include/llvm/Object/ELF.h

653

@namhyung, I believe @MaskRay was suggesting using the comment he's posted instead of "no section string table", as an improvement. Please update.

778–780

Having the "symmetry" has the potential to actually cause confusion, since the LHS isn't negated unlike the RHS.

llvm/test/Object/objdump-no-sectionheaders.test

2

22

I think I see a potential problem here, which could explain the problem you're seeing: you haven't specified any contents or explicit offsets of any of these program headers, which may mean that even though they have addresses specified, they'll start at the same offset within the file. This might result in something bad such as only the one section appearing in the output, although I don't know the llvm-objdump code well enough to be certain.

[namhyung]

llvm/include/llvm/Object/ELF.h

653

Oh, I overlooked that, sorry. Will update.

llvm/test/Object/objdump-no-sectionheaders.test

2

ok

22

Will check with contents and offsets.

[namhyung]

v3 updates. But the test still fails.

• add content and offset to phdr in the test
• update comments for fake section strings
• remove unnecessary paranthesis

[jhenderson]

llvm/test/Object/objdump-no-sectionheaders.test

2

I misread this sentence. Here's a better improvement.

22

The Offset will physically impact the size of the file on disk, meaning this generated YAML file will be huge. Instead, use offset values around 0x100 - 0x200, which are significantly smaller (and adjust the addresses to line up too, I recommend) (you can't use 0 because of the ELF header and Program Header table). Also, the sizes of the individual blobs of data can be only a couple of bytes for the test purposes to be satisfied.

[namhyung]

v4 updates. But the test still fails..

• update comment in the test
• use smaller offsets for the test input

[jhenderson]

Have you tried using a debugger to investigate the test failure?

[namhyung]

I suspect the StringTableBuilder does something wrong, or my understanding is incorrect. I checked with a debugger that the sh_name offset set correctly, but the first string was "PT_LOAD#3" instead of "PT_LOAD#1". Investigating...

[namhyung]

Sorry for the delay. I think I found the reason - it passed a stack-allocate string to the string table builder which keeps the cached string ref. So each section name pointed to the same place in stack and got overwritten. Will add a vector to hold the local strings.

[namhyung]

v5 update. Now it passes the test!

• use a vector for local section name strings

[MaskRay]

LGTM.

llvm/include/llvm/Object/ELF.h

777–778

for (auto [Idx, Phdr] : llvm::enumerate(*PhdrsOrErr))

[jhenderson]

llvm/include/llvm/Object/ELF.h

789

I'm pretty sure this isn't going to work as intended in many cases: you create a StringRef to a std::string temporary, and stick that in SecNames. It's likely that the StringRef will be pointing at invalid memory immediately after its construction (it's likely just working because nothing has scribbled over the data in memory yet). Could you not just populate SecNames with the result of the "PT_LOAD#" + std::to_string(Idx) expression directly?

If the StringTableBuilder contains StringRefs to the strings, using a std::vector will potentially invalidate them, if it needs to increase its capacity, right? We probably want a different strategy here. Probably simplest is simply to give the vector an initial capacity equal to the total number of program headers. Given that there are rarely that many, this shouldn't have any significant impact on performance.

[MaskRay]

llvm/include/llvm/Object/ELF.h

789

Thanks for catching this. It is a code smell but it actually works: "PT_LOAD#" + std::to_string(Idx) returns a std::string. It works as the the lifetime of the temporary object std::string applies to the full-expression.

The following should be better:

SmallVector<std::string, 0> SecNames; 
...
SecNames.push_back(("PT_LOAD#" + Twine(Idx)).str());

// An alternative is ("PT_LOAD#" + Twine(Idx)).toVector(Name); but a local `SmallString` variable is needed

[namhyung]

SmallVector<std::string, 0> SecNames; 
...
SecNames.push_back(("PT_LOAD#" + Twine(Idx)).str());

This makes the test failing again.

[jhenderson]

In D131290#3771385, @namhyung wrote:

SmallVector<std::string, 0> SecNames; 
...
SecNames.push_back(("PT_LOAD#" + Twine(Idx)).str());

This makes the test failing again.

Please see the second half of my previous comment re. vector resizing (can apply to SmallVector too potentially). I would expect that to be a problem.

[MaskRay]

In D131290#3771459, @jhenderson wrote:

In D131290#3771385, @namhyung wrote:

SmallVector<std::string, 0> SecNames; 
...
SecNames.push_back(("PT_LOAD#" + Twine(Idx)).str());

This makes the test failing again.

Please see the second half of my previous comment re. vector resizing (can apply to SmallVector too potentially). I would expect that to be a problem.

I think we should switch to what I originally suggested, drop StringTableBuilder and use a string with \0 terminators:)

[jhenderson]

In D131290#3771470, @MaskRay wrote:

In D131290#3771459, @jhenderson wrote:

In D131290#3771385, @namhyung wrote:

SmallVector<std::string, 0> SecNames; 
...
SecNames.push_back(("PT_LOAD#" + Twine(Idx)).str());

This makes the test failing again.

Please see the second half of my previous comment re. vector resizing (can apply to SmallVector too potentially). I would expect that to be a problem.

I think we should switch to what I originally suggested, drop StringTableBuilder and use a string with \0 terminators:)

I'm actually inclined to agree - the "simplification" is turning out to be more complicated than I expected.

[namhyung]

In D131290#3771526, @jhenderson wrote:

In D131290#3771470, @MaskRay wrote:

In D131290#3771459, @jhenderson wrote:

In D131290#3771385, @namhyung wrote:

SmallVector<std::string, 0> SecNames; 
...
SecNames.push_back(("PT_LOAD#" + Twine(Idx)).str());

This makes the test failing again.

Please see the second half of my previous comment re. vector resizing (can apply to SmallVector too potentially). I would expect that to be a problem.

I think we should switch to what I originally suggested, drop StringTableBuilder and use a string with \0 terminators:)

I'm actually inclined to agree - the "simplification" is turning out to be more complicated than I expected.

Let me try one more time before going back. I think this failure came from the std::string implementation. From a quick glance it seems to store small strings inside the object, not in the external allocation. Then a later move due to vector resize will invalidate earlier references. But SmallString seems to store the data always in the external allocation so it'd be fine after move. Actually it passes the test when I change it like below

-  SmallVector<std::string, 0> SecNames;
+  SmallVector<SmallString<10>, 0> SecNames;

Let me know if I missed something.

[jhenderson]

I don't think you can rely on the behaviour of what happens to the strings when vectors are resized - I don't believe there's anything in the C++ standard or the existing type's definition that will guarantee the behaviour will remain the same. Also, your explanation sounds reversed to me, although I haven't looked at the implementation: I'd expecting SmallString to have a stack-allocated array for storing strings when the size is small, but std::string to always use the heap. In fact, on some implementations, I wouldn't be surprised if std::string also uses stack-allocation for short strings (I recall running into a problem very similar to yours along the same lines recently due to this sort of thing). Either way, it cannot be relied upon to keep your references valid.

[namhyung]

llvm/include/llvm/Object/ELF.h

789

I'm pretty sure this isn't going to work as intended in many cases: you create a StringRef to a std::string temporary, and stick that in SecNames. It's likely that the StringRef will be pointing at invalid memory immediately after its construction (it's likely just working because nothing has scribbled over the data in memory yet).

I don't think so. The StringRef was needed because the constructor of SmallString takes it only. I wanted to pass std::string directly but it's not supported IIUC. After it creates a SmallString, I don't use the StringRef for a temp string anymore. Instead it uses the SmallString data in the vector, that's why I passed SecNames.back(). I believe it's safe to refer the data even after vector is resized.

[namhyung]

In D131290#3773833, @jhenderson wrote:

I don't think you can rely on the behaviour of what happens to the strings when vectors are resized - I don't believe there's anything in the C++ standard or the existing type's definition that will guarantee the behaviour will remain the same.

Fair enough. Let's go back to use a string directly. :)

[namhyung]

v6 update.

• do not use StringTableBuilder

[jhenderson]

llvm/include/llvm/Object/ELF.h

778–780

Do you actually need this?

[namhyung]

llvm/include/llvm/Object/ELF.h

778–780

Yep, Idx is a part of the section name.

[jhenderson]

llvm/include/llvm/Object/ELF.h

778–780

I wasn't referring to MaskRay's suggestion. I was referring to the line that's actually highlighted with this comment (i.e. line 778 in this diff).

[namhyung]

llvm/include/llvm/Object/ELF.h

778–780

Oh, I see. Maybe it's not needed but I think it's a good practice to maintain offset 0 as an empty string. Do you want me to remove it?

[MaskRay]

llvm/include/llvm/Object/ELF.h

778–780

FakeSectionStrings += '\0' or use push_back.

llvm/test/Object/objdump-no-sectionheaders.test

1–2

Nit: "This test " can be omitted. Use an imperative sentence.

[jhenderson]

LGTM, with @MaskRay's nits addressed.

llvm/include/llvm/Object/ELF.h

778–780

Having come back to this a day later, I think it's wise to keep it, as we don't want to risk there being code that assumes sh_name == '\0' means no name.

llvm/test/Object/objdump-no-sectionheaders.test

1–2

Concretely: "Check llvm-objdump -h ..."

[namhyung]

v7 update.

• use SmallString::operator+=()
• update comment in the test

[namhyung]

Thank you both for the review! Please commit with 'Namhyung Kim <[e-mail]>'.

[jhenderson]

In D131290#3776478, @namhyung wrote:

Thank you both for the review! Please commit with 'Namhyung Kim <[e-mail]>'.

All submitted. Please keep an eye out for any bot failures.

llvm/include/llvm/Object/ELF.h

 //===- ELF.h - ELF object file implementation -------------------*- C++ -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
 // This file declares the ELFFile template class.
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_OBJECT_ELF_H
 #define LLVM_OBJECT_ELF_H
 
 #include "llvm/ADT/ArrayRef.h"
+#include "llvm/ADT/SmallString.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/BinaryFormat/ELF.h"
 #include "llvm/Object/ELFTypes.h"
 #include "llvm/Object/Error.h"
 #include "llvm/Support/Endian.h"
 #include "llvm/Support/Error.h"
 #include <cassert>
public:
   const uint8_t *base() const { return Buf.bytes_begin(); }
   const uint8_t *end() const { return base() + getBufSize(); }
 
   size_t getBufSize() const { return Buf.size(); }
 
 private:
   StringRef Buf;
   std::vector<Elf_Shdr> FakeSections;
+  SmallString<0> FakeSectionStrings;
 
   ELFFile(StringRef Object);
 
 public:
   const Elf_Ehdr &getHeader() const {
     return *reinterpret_cast<const Elf_Ehdr *>(base());
   }
 
if (Index == ELF::SHN_XINDEX) {
     // header at index 0.
     if (Sections.empty())
       return createError(
           "e_shstrndx == SHN_XINDEX, but the section header table is empty");
 
     Index = Sections[0].sh_link;
   }
 
-  if (!Index) // no section string table.
-    return "";
+  // There is no section name string table. Return FakeSectionStrings which
+  // is non-empty if we have created fake sections.
+  if (!Index)
+    return FakeSectionStrings;
+
   if (Index >= Sections.size())
     return createError("section header string table index " + Twine(Index) +
                        " does not exist");
   return getStringTable(Sections[Index], WarnHandler);
 }
 
 /// This function finds the number of dynamic symbols using a GNU hash table.
 ///

 /// disassemble objects without a section header table (e.g. ET_CORE objects
 /// analyzed by linux perf or ET_EXEC with llvm-strip --strip-sections).
 template <class ELFT> void ELFFile<ELFT>::createFakeSections() {
   if (!FakeSections.empty())
     return;
   auto PhdrsOrErr = program_headers();
   if (!PhdrsOrErr)
     return;
 
-  for (auto Phdr : *PhdrsOrErr) {
-    if (!(Phdr.p_type & ELF::PT_LOAD) || !(Phdr.p_flags & ELF::PF_X))
+  FakeSectionStrings += '\0';
+  for (auto [Idx, Phdr] : llvm::enumerate(*PhdrsOrErr)) {
+    if (Phdr.p_type != ELF::PT_LOAD || !(Phdr.p_flags & ELF::PF_X))
       continue;
     Elf_Shdr FakeShdr = {};
     FakeShdr.sh_type = ELF::SHT_PROGBITS;
     FakeShdr.sh_flags = ELF::SHF_ALLOC | ELF::SHF_EXECINSTR;
     FakeShdr.sh_addr = Phdr.p_vaddr;
     FakeShdr.sh_size = Phdr.p_memsz;
     FakeShdr.sh_offset = Phdr.p_offset;
+    // Create a section name based on the p_type and index.
+    FakeShdr.sh_name = FakeSectionStrings.size();
+    FakeSectionStrings += ("PT_LOAD#" + Twine(Idx)).str();
+    FakeSectionStrings += '\0';
     FakeSections.push_back(FakeShdr);
   }
 }
 
 template <class ELFT>
 Expected<typename ELFT::ShdrRange> ELFFile<ELFT>::sections() const {
   const uintX_t SectionTableOffset = getHeader().e_shoff;
   if (SectionTableOffset == 0) {

llvm/test/Object/objdump-no-sectionheaders.test

-; RUN: llvm-objdump -h %p/Inputs/no-sections.elf-x86-64 \
-; RUN:              | FileCheck %s
+## Check llvm-objdump -h can handle ELF files without section info.
+## Only PT_LOAD segments with the PF_X flag will be displayed as fake sections.
 
-; CHECK:      Sections:
-; CHECK-NEXT: Idx Name          Size     VMA              Type
-; CHECK-NEXT:   0               000006ec 0000000000400000 TEXT
-; CHECK-NEXT:   1               00000000 0000000000000000 TEXT
-; CHECK-NOT:  {{.}}
+# RUN: yaml2obj %s -o %t
+# RUN: llvm-objdump -h %t | FileCheck %s
+
+# CHECK:      Sections:
+# CHECK-NEXT: Idx Name          Size     VMA              Type
+# CHECK-NEXT:   0 PT_LOAD#1     00000100 0000000000400000 TEXT
+# CHECK-NEXT:   1 PT_LOAD#3     00000200 0000000000600400 TEXT
+# CHECK-NOT:  {{.}}
+
+!ELF
+FileHeader:
+  Class:           ELFCLASS64
+  Data:            ELFDATA2LSB
+  Type:            ET_CORE
+  Machine:         EM_X86_64
+Sections:
+  - Type:            SectionHeaderTable
+    NoHeaders:       true
+  - Type:            Fill
+    Name:            code1
+    Pattern:         "cc"
+    Size:            0x100
+    Offset:          0x200
+  - Type:            Fill
+    Name:            data1
+    Pattern:         "aa55"
+    Size:            0x100
+    Offset:          0x300
+  - Type:            Fill
+    Name:            code2
+    Pattern:         "ff"
+    Size:            0x200
+    Offset:          0x400
+ProgramHeaders:
+  - Type:            PT_PHDR
+    Flags:           [ PF_X ]
+    VAddr:           0x400000
+    MemSize:         0x100
+  - Type:            PT_LOAD
+    Flags:           [ PF_X ]
+    VAddr:           0x400000
+    MemSize:         0x100
+    FirstSec:        code1
+    LastSec:         code1
+  - Type:            PT_LOAD
+    Flags:           [ PF_R ]
+    VAddr:           0x500300
+    MemSize:         0x100
+    FirstSec:        data1
+    LastSec:         data1
+  - Type:            PT_LOAD
+    Flags:           [ PF_R, PF_X ]
+    VAddr:           0x600400
+    MemSize:         0x200
+    FirstSec:        code2
+    LastSec:         code2

llvm/test/tools/llvm-objdump/X86/disassemble-no-section.test

 ## This test checks -d disassembles an ELF file without section headers.
 ## Such files include kcore files extracted by linux perf tools, or
 ## executables with section headers stripped by e.g.
 ## llvm-strip --strip-sections.
 
 # RUN: yaml2obj %s -o %t
 # RUN: llvm-objdump -d %t | FileCheck %s
 
-# CHECK:       Disassembly of section :
+# CHECK:       Disassembly of section PT_LOAD#0:
 # CHECK-EMPTY:
-# CHECK-NEXT:  <>:
+# CHECK-NEXT:  <PT_LOAD#0>:
 # CHECK-NEXT:  55                    pushq   %rbp
 # CHECK-NEXT:  48 89 e5              movq    %rsp, %rbp
 # CHECK-NEXT:  0f 1f 40 00           nopl    (%rax)
 # CHECK-NEXT:  5d                    popq    %rbp
 # CHECK-NEXT:  c3                    retq
 
 ## Check disassembly with an address range.
 # RUN: llvm-objdump -d --start-address=0xffffffff00000000 \
 # RUN:   --stop-address=0xffffffff00000004 %t 2>&1 | \
 # RUN:   FileCheck %s --check-prefix RANGE
 
 # RANGE:       no section overlaps the range
 # RANGE-EMPTY:
-# RANGE-NEXT:  Disassembly of section :
+# RANGE-NEXT:  Disassembly of section PT_LOAD#0:
 # RANGE-EMPTY:
-# RANGE-NEXT:  <>:
+# RANGE-NEXT:  <PT_LOAD#0>:
 # RANGE-NEXT:  55                    pushq   %rbp
 # RANGE-NEXT:  48 89 e5              movq    %rsp, %rbp
 # RANGE-EMPTY:
 
 !ELF
 FileHeader:
   Class:           ELFCLASS64
   Data:            ELFDATA2LSB

llvm/tools/llvm-objdump/llvm-objdump.cpp

static size_t getMaxSectionNameWidth(const ObjectFile &Obj) {
   for (const SectionRef &Section : ToolSectionFilter(Obj)) {
     StringRef Name = unwrapOrError(Section.getName(), Obj.getFileName());
     MaxWidth = std::max(MaxWidth, Name.size());
   }
   return MaxWidth;
 }
 
 void objdump::printSectionHeaders(ObjectFile &Obj) {
+  if (Obj.isELF() && Obj.sections().empty())
+    createFakeELFSections(Obj);
+
   size_t NameWidth = getMaxSectionNameWidth(Obj);
   size_t AddressWidth = 2 * Obj.getBytesInAddress();
   bool HasLMAColumn = shouldDisplayLMA(Obj);
   outs() << "\nSections:\n";
   if (HasLMAColumn)
     outs() << "Idx " << left_justify("Name", NameWidth) << " Size     "
            << left_justify("VMA", AddressWidth) << " "
            << left_justify("LMA", AddressWidth) << " Type\n";
   else
     outs() << "Idx " << left_justify("Name", NameWidth) << " Size     "
            << left_justify("VMA", AddressWidth) << " Type\n";
 
-  if (Obj.isELF() && Obj.sections().empty())
-    createFakeELFSections(Obj);
-
   uint64_t Idx;
   for (const SectionRef &Section : ToolSectionFilter(Obj, &Idx)) {
     StringRef Name = unwrapOrError(Section.getName(), Obj.getFileName());
     uint64_t VMA = Section.getAddress();
     if (shouldAdjustVA(Section))
       VMA += AdjustVMA;
 
     uint64_t Size = Section.getSize();