ifunc resolver parameters

When ifunc support was added to Clang with r265917
http://llvm.org/viewvc/llvm-project?rev=265917&view=rev it did not
allow resolvers to take parameters. It appears GCC documents ifunc
resolvers as taking no parameters, but GCC in fact allows it.

In FreeBSD our rtld passes machine-dependent CPU ID information to
ifunc resolvers, and we had to remove the prohibition:
[base] Revision 339019.

In the glibc world, as far as I can tell it in fact passes dl_hwcap on
several architectures (but not on x86 or x86_64).

I believe GCC's documentation is simply incorrect. Is there any reason
not to just remove this check and error case from Clang?

When I implemented ifunc I followed GCC documentation and was not aware of ifunc resolvers parameter. But if it is so much platform dependent and undocumented in GCC, I’m doubt if it should be supported in clang.

(+echristo in case he has some thoughts here)

I've raised 87482 – Clarify behaviour of resolvers with parameters in for __attribute__((ifunc)) to see
if I can get some clarification about the intent of the documentation.

On some targets like Arm and AArch64 the glibc dynamic loader will
pass a parameter HWCAPS to the ifunc resolver function that the
resolver can use choose the function, whereas I believe it does not on
x86_64 where it is common for the ifunc resolver to check the cpu id
directly to get the same information.

In the Arm and AArch64 ports of glibc ifunc resolvers with parameters
are used but __attribute__((ifunc("resolver"))) isn't used. A somewhat
lower level method using inline assembly is used instead: asm (".type
func1, %gnu_indirect_function");

My personal view is that the error would be useful in the context of
(X86_64, glibc) but not in the case of (AArch64, glibc) or (Arm,
glibc), it is also possible that FreeBSD have chosen to pass a
parameter even in X86_64. So I think that the error message, if it is
there at all, should be conditional on a Target.

Peter

Yea, definitely clang shouldn’t error out on correct usage (which taking a parameter is on some platforms).

What the type should be depends on the dynamic loader implementation being used. Glibc variously passes uint64_t, unsigned long, int, or nothing. (from a grep for “elf_irel”). Other linux libc implementations could well do other things. Certainly other OS’s libc do. It doesn’t really seem worthwhile to try to encode all the possible function signatures required here for the different targets…I’d suggest that simply removing the error message would be best.

I have review D52703 open for removing the error message.