Preemptable linkage

[philipc]

For defined symbols, faerie always references the section symbol in ELF relocations:

faerie/src/elf.rs

Line 676 in dcb671a

(to_shndx - 3) + self.special_symbols.len()

I think this means that these relocations can never be preempted. However, my understanding is that normally all global symbols are preemptable, regardless of whether they are weak.

e.g. for the following C code:

static int local() { return 1; }
int global() { return 1; }
__attribute__((weak)) int weak() { return 1; }

int bar(int x) {
	return local() + global() + weak();
}

I get these relocations (compiled with gcc -fPIC -ffunction-sections):

  Offset          Info           Type           Sym. Value    Sym. Name + Addend
000000000012  000500000002 R_X86_64_PC32     0000000000000000 .text.local - 4
00000000001e  000d00000004 R_X86_64_PLT32    0000000000000000 global - 4
00000000002a  000e00000004 R_X86_64_PLT32    0000000000000000 weak - 4

To complicate this, cranelift has Linkage::Preemptible, but this is translated to weak for the faerie backend, which seems to be conflating weak and preemptable.

Is my understanding of all this correct?

cc @pchickey since I think you did the weak support?

[pchickey]

I don't follow the C code example. Is faerie not capable of emitting those relocations? How is the difference related to the symbol table?

I don't think I understand the distinction between preemptable and weak. My understanding was that @sunfishcode selected those variants because the two were synonyms.

[philipc]

For the C code example, note that the global and weak relocations reference the symbol, not the section (e.g. global and not .text.global). faerie can't emit them because it always uses the section (and often using the section is right, such as for the local symbol).

For preemptable vs weak, maybe I am reading too much into the name. Have a look at https://llvm.org/docs/LangRef.html#runtime-preemption-specifiers. It would be good if @sunfishcode could clarify the intent here.

[pchickey]

Thanks, I missed the distinction between referencing a symbol and a section. I agree that this is something we should correct.

[sunfishcode]

@philipc I believe you are correct, and we should fix this.

[philipc]

@sunfishcode Is cranelift's Linkage::Preemptible meant to be the same as weak, or do we need to add another variant for it?