On 16 November 2017 at 21:37, Sami Tolvanen [off-list ref] wrote:

On Thu, Nov 16, 2017 at 04:34:03PM +0000, Ard Biesheuvel wrote:

quoted

You still have not explained to us how GOLD avoids the erratum.

Sorry, I didn't realize you were asking that. If gold spots erratum
sequences, looks like it creates stubs to break them up:

  https://sourceware.org/git/?p=binutils-gdb.git;a=blob;f=gold/aarch64.cc#l8396

It also attempts to optimize the code by replacing adrps in these
sequences with adr where possible, but otherwise doesn't appear to
touch them:

  https://sourceware.org/git/?p=binutils-gdb.git;a=blob;f=gold/aarch64.cc#l2053

OK, so my concern here is that this code probably only operates on
fully linked binaries, and not partially linked object files like
kernel modules. The same applies to ld.bfd, which is why we need to
use the large module instead of a code model that may emit adrp
instructions..

What is preventing us from using the large model with clang? I know it
uses movk/movz pairs rather than literals, but this shouldn't matter
for modules, given that we support static ELF relocations.