Vague linkage

In C++, inline functions, template instantiations, and a few other things can be defined in multiple object files but need deduplication at link time. In the dark ages the functionality was implemented by weak definitions: the linker does not report duplicate definition errors and resolves the references to the first definition. The downside is that unneeded copies remained in the linked image.

COMDAT

In the Microsoft PE file format, COMDAT was invented. It has two major features: (1) enabling deduplication of a set of sections (2) allowing a set of sections to be retained or discarded as a unit.

The section flag (IMAGE_SCN_LNK_COMDAT) marks a section COMDAT and enables deduplication on a per-section basis.

The first symbol that has the section value of the COMDAT section must be the section symbol. The second symbol is called "the COMDAT symbol" and is used by the linker in conjunction with the Selection field in its section definition auxiliary record. The Selection field describes how the associated section should be deduplicated.

IMAGE_COMDAT_SELECT_ANY is the most common selection value. "Any section that defines the same COMDAT symbol can be linked; the rest are removed." link.exe and ld.lld just pick the first seen section. Similar selection kinds include: IMAGE_COMDAT_SELECT_SAME_SIZE, IMAGE_COMDAT_SELECT_EXACT_MATCH, and IMAGE_COMDAT_SELECT_LARGEST.

IMAGE_COMDAT_SELECT_ASSOCIATIVE makes a section linked to another section. The interrelated sections will be retained or discarded as a unit. Conceptually there is a leader and an arbitrary number of followers. We will see that in ELF section groups the sections are equal.

If a text section needs a data section and deduplication is needed for both sections, you have two choices:

• Use two COMDAT symbols. There is the drawback that deduplication happens independently for the interconnected sections.
• Make the data section link to the text section via IMAGE_COMDAT_SELECT_ASSOCIATIVE. Whether an IMAGE_COMDAT_SELECT_ASSOCIATIVE section is retained is dependent on its referenced section.

IMAGE_COMDAT_SELECT_NODUPLICATES is interesting. The implementation just does not perform deduplication. If a section defines an external symbol and the section is duplicated in another object file, there will be a duplicate definition linker error.

link.exe limitation

There is a limitation: MSVC link.exe may report a duplicate symbol error (error LNK2005) for an external symbol defined in a IMAGE_COMDAT_SELECT_ASSOCIATIVE section, even if it would be discarded after handling the leader symbol.

ELF section groups

In the GNU world, .gnu.linkonce. was invented to deduplicate groups with just one member. .gnu.linkonce. has been long obsoleted in favor of section groups but the usage has been daunting until 2020. Adhemerval Zanella removed the last live glibc use case for .gnu.linkonce. BZ #20543.

Cary Coutant et al implemented section groups and GRP_COMDAT on HP-UX before pushing the idea to the generic ABI committee. The design was a generalization of PE COMDAT. You can find the document IA-64 gABI Issue 72: COMDAT on the Internet. It lists some design consideration.

Some sections occur in interrelated groups. For example, an out-of-line definition of an inline function might require, in addition to the section containing its executable instructions, a read-only data section containing literals referenced, one or more debugging information sections and other informational sections. Furthermore, there may be internal references among these sections that would not make sense if one of the sections were removed or replaced by a duplicate from another object. Therefore, such groups must be included or omitted from the linked object as a unit. A section cannot be a member of more than one group.

You may want to read the description before reading the following paragraphs.

According to "such groups must be included or omitted from the linked object as a unit", a linker's garbage collection feature must retain or discard the sections as a unit. This property can be leveraged by interrelated metadata sections.

Say __llvm_prf_data and __llvm_prf_cnts are used as interrelated metadata sections describing the text section .text.a. __llvm_prf_data has a relocation to __llvm_prf_cnts. .text.a and other text sections reference __llvm_prf_cnts. Without the "as a unit" semantics, ld --gc-sections may retain __llvm_prf_cnts while discarding __llvm_prf_data, since __llvm_prf_data is not referenced by a live section. The "as a unit" semantics allow __llvm_prf_data to be retained due to its interrelation with __llvm_prf_cnts.

A section group can contain more than 2 members. E.g. if the function in .text.a uses value profiling, there will be a third metadata section __llvm_prf_vnds, which is not referenced. The "as a unit" semantics allow __llvm_prf_vnds to be retained if __llvm_prf_data or __llvm_prf_cnts is retained.

The section data of a SHT_GROUP section is an array of Elf32_Word entries. The first entry is a flag word. The remaining entries are a sequence of section header indices.

GRP_COMDAT

The most common section group flag is GRP_COMDAT, which makes the member sections similar to COMDAT in Microsoft PE file format, but can apply to multiple sections. (The committee borrowed the name "COMDAT" from PE.)

GRP_COMDAT - This is a COMDAT group. It may duplicate another COMDAT group in another object file, where duplication is defined as having the same group signature. In such cases, only one of the duplicate groups may be retained by the linker, and the members of the remaining groups must be discarded.

The signature symbol provides a unique name for deduplication. The binding (local/global status) of the group signature is not part of the equation. This is not clear in the specification and has been clarified by GRP_COMDAT group with STB_LOCAL signature.

In D106782, I changed llvm-objcopy to clear the GRP_COMDAT bit if the signature symbol becomes local. --localize-hidden and --keep-global-symbol are sometimes used to localize symbols. If the signature symbol of a section group is localized, the intention is likely to localize the section group and suppress deduplication. GNU objcopy feature request: PR27931.

I want to highlight one thing GCC does (and Clang inherits) for backward compatibility: definitions relative to a COMDAT section are usually STB_WEAK instead of STB_GLOBAL. The idea is that an old toolchain which does not recognize COMDAT groups can still operate correctly, just in a degraded manner.

The section group flag can be 0: no signature based deduplication should happen.

LLVM IR

In LLVM IR, comdat provides access to object file COMDAT/section group functionality which represents interrelated sections. The syntax is closer to ELF section groups in spirit. I recently added clarification that "a comdat must be included or omitted as a unit."

For example, the following IR defines a comdat named foo and its two members foo and bar.

1
2
3

$foo = comdat any
@foo = global i32 2, comdat
@bar = global i32 3, comdat($foo)

In PE/COFF, the comdat must contain a member which has the same name as the group. The member (leader) will encode the IR selection kind and the other members will use the IMAGE_COMDAT_SELECT_ASSOCIATIVE selection.

llvm.global_ctors and llvm.global_dtors

On ELF, llvm.global_ctors elements are usually placed in .init_array and .init_array.* sections.

If the third field is non-null, and points to a global variable or function, the initializer function will only run if the associated data from the current module is not discarded. On ELF the referenced global variable or function must be in a comdat.

The third field was initially added for PE/COFF so that linker garbage collection can discard an initialization section if its referenced text section can be discarded. If the referenced text section is in a comdat any, the initialization section can be deduplicated with other translation units.

On ELF, the third field is implemented as a section group to enable deduplication.

As an example, -fsanitize-coverage= instrumentation generates:

1
2
3
4
5
6
7
8
9

$sancov.module_ctor_trace_pc_guard = comdat any

@llvm.global_ctors = appending global [1 x { i32, void ()*, i8* }] [{ i32, void ()*, i8* } { i32 2, void ()* @sancov.module_ctor_trace_pc_guard, i8* bitcast (void ()* @sancov.module_ctor_trace_pc_guard to i8*) }]
@llvm.used = appending global [1 x i8*] [i8* bitcast (void ()* @sancov.module_ctor_trace_pc_guard to i8*)], section "llvm.metadata"

define internal void @sancov.module_ctor_trace_pc_guard() #1 comdat {
  call void @__sanitizer_cov_trace_pc_guard_init(i32* @__start___sancov_guards, i32* @__stop___sancov_guards)
  ret void
}

The module constructor sancov.module_ctor_trace_pc_guard is in an IR comdat. The llvm.global_ctors element compiles to a .init_array.2 in the same section group as .text.sancov.module_ctor_trace_pc_guard.

1
2
3
4
5
6
7

% llvm-readelf -g a.o
COMDAT group section [    6] `.group' [sancov.module_ctor_trace_pc_guard] contains 4 sections:
   [Index]    Name
   [    7]   .text.sancov.module_ctor_trace_pc_guard
   [    8]   .rela.text.sancov.module_ctor_trace_pc_guard
   [   11]   .init_array.2
   [   12]   .rela.init_array.2

.init_array.2 has the type SHT_INIT_ARRAY. Such a section within a group is unusual. To the best of my knowledge, such sanitizer usage was not leveraged by GCC/Clang before.

Before Clang 13.0.0, the SHT_INIT_ARRAY member in a group was a GC root. It ensured the module constructor could not be discarded by linker GC. This is considered as an abuse (see GC-ability of SHT_INIT_ARRAY). I placed the module constructor in llvm.used so that .text.sancov.module_ctor_trace_pc_guard has the SHF_GNU_RETAIN flag.

ld.lld doesn't treat SHT_INIT_ARRAY sections with the SHF_LINK_ORDER flag as GC roots. Linkers do not treat SHT_NOTE sections within a group as GC roots. For consistency, I made SHT_INIT_ARRAY in a section group GCable for LLD 13.0.0.

WebAssembly

comdat nodeduplicate is not implemented. I filed a feature request https://bugs.llvm.org/show_bug.cgi?id=50531.

It strikes me that WebAssembly need WASM_COMDAT_DATA, WASM_COMDAT_FUNCTION, and WASM_COMDAT_SECTION to encode COMDAT. It is difficult for an elf to understand the complexity:)

C inline

-fgnu89-inline inline and C99 inline are different from C++ inline. Non-static non-extern inline has weird semantics. I think it may be because the designers wanted to control the defining translation unit and avoid vague linkage. The choices did avoid the need for COMDAT but made the non-static non-extern inline inconvenient to use. In practice I think most projects avoid it.

Case study - PGO/coverage

A function and its data/counter variables must use separate IR comdat. (a) Instrumentation runs before inlining. More functions may reference the function's data/counter. Using one comdat can lead to dangling references.

2. Using separate comdat can allow benign ODR violation where the prevailing copy is from a non-instrumented module.

Make FunctionPointer relative

Using a private alias to the function does not allow this scenario: both a non-instrumented TU and a instrumented TU have a function in a comdat any. If the non-instrumented TU is prevailing, a relocation referencing the discarded private alias will cause a linker error.

Q: How to handle an available_externally function? (e.g. extern __inline __attribute__((gnu_inline)) int stat(const char *path, struct stat *statbuf))

An alias cannot reference an available_externally function.