lld does not seem to support __start and __stop

[EdSchouten]

Bugzilla Link	22906
Resolution	FIXED
Resolved on	Feb 25, 2016 02:55
Version	unspecified
OS	All
Blocks	#23588
CC	@emaste

Extended Description

GNU ld supports the magic __start and __stop symbols as markers for the beginning/end of sections. For example:

int var1 attribute((section("foo")));
int var2 attribute((section("foo")));

extern int __start_foo[0];
extern int __stop_foo[0];

GNU ld automatically adds these symbols if undefined. They will point to the beginning/end of the section "foo". Couple of important things to keep in mind when implementing this:

• Don't let --gc-sections reap these sections if __start_foo and __stop_foo are being used.
• Don't inject these symbols if the section itself does not exist.

GNU ld also doesn't do this.

[llvmbot]

What happens when there are multiple sections with the same name?

Is this also supported by gold or other elf linkers?

[EdSchouten]

Hi Rafael,

GNU Gold also supports this construct (see Gold's Layout::define_section_symbols()).

The linker is supposed to ensure that all sections with the same name are placed in the resulting binary contiguously, right? Is it even possible to generate a binary that has two sections with the same name?

[llvmbot]

Hi Rafael,

GNU Gold also supports this construct (see Gold's
Layout::define_section_symbols()).

The linker is supposed to ensure that all sections with the same name are
placed in the resulting binary contiguously, right? Is it even possible to
generate a binary that has two sections with the same name?

Oh, the symbols are for the concatenated output section. That makes sense. For some reason I read the description as being for each .o.

[emaste]

Related are the special symbols _etext, _edata, _end

[emaste]

http://llvm.org/viewvc/llvm-project?rev=241341&view=rev

[emaste]

ELF2 implementation http://llvm.org/viewvc/llvm-project?view=revision&revision=250432

[EdSchouten]

Reopening this, as it has not been implemented completely.

The problem with the existing implementation is that __start and __stop symbols are only added after reachability of sections has been computed. What I mean is that __start and __stop cannot be used to mark sections as live. They will get removed if --gc-sections is being used.

This is different from GNU Binutils, were a reference to __start or __stop is sufficient to prevent the section from being garbage collected.

[llvmbot]

This should help http://reviews.llvm.org/D17502

[llvmbot]

Fix is available: r261840.

But commit message contains incorrect info: behavior currently is not the same as in GNU binutils. While binutils takes in account references to _start/_stop symbols, we just do not reclaim sections if their section names are valid C identifiers.

[emaste]

mentioned in issue #23588