Julia Evans

Day 40: 12 things I learned today about linkers.

I read 11 parts of this series on linkers today. I also wrote an epic blog post, but here is the tl;dr version (trying to synthesize…). This is all about ELF. I use “ELF file” and “object file” interchangeably.

In no particular order:

  1. To inspect an ELF object file, you can use objdump, readelf and/or nm.
  2. Executable files have segments and sections. Each segment has many sections. The operating system looks at the segments, not the sections. Read/Write/Execute permissions are controlled per segment, not per section. [Part 8]
  3. ELF symbols have types! And different visibility options! [Part 5]
  4. The linker knows about threading, and does optimizations to make threading easier. [Part 7]
  5. An object file can define two symbols with the same name and different symbols, for backwards compatibility. [Part 9]
  6. Those .a files? Those are just collections of .o object files, and they’re called “archives”! [Part 11]
  7. Linkers can work in parallel to some extent. [Part 10]
  8. Linkers actually have to do fairly complicated stuff to allow the code in a shared library to be shared between different programs and save memory. [Part 6] for memory savings, [Part 4] for the PLT/GOT
  9. There’s more than one way to link a shared library, and the choices you make affect how quickly it loads [Part 4]
  10. In the Mach-O executable format you can have assembly code for differerent architectures in the same executable. Nuts. And there’s FatELF that extends ELF to do the same thing. (edit: and isn’t being developed anymore)
  11. Every .o file has a “relocation table” listing every single reference to a symbol that the linker will need to update, and how it will need to update it. [Part 2]
  12. If you’re making a speed comparison for a linker, you might compare it to cat. [Part 1]

I’m curious about these ELF symbol versions – they sound kind of like polymorphism to me, and I’m wondering why people use symbol name mangling to implement polymorphism instead of symbol versions. Probably very good reasons!

Day 39: Writing malloc wrong, for fun Day 40: Linkers are amazing.