On Sat, Apr 22, 2023 at 7:20 PM Samuel Thibault <samuel.thiba...@gnu.org> wrote: > But what about the memcpy calls from the RPC stubs used by the init > code?
Ah, I see that we do actually call more than I thought before _start1. In particular, we initialize the dtable (so the ctty stuff...) and get our pid/pgrp from the proc server. We should probably do this later, at time __libc_init_first probably. We want to do as little as possible before _start1; we must get our argv and env, so that must happen before _start1, and to do that we need to set up __init_tcbhead, but the rest we can do later. Guess I still like my GOT prefilling approach better then. It's not that big of a hack, it's just two lines of assembly (we can drop stpcpy, it's not used during early startup) and it makes memcpy just work from anywhere (any RPCs, any other code) without any additional trickery at use sites or in the build system. And all the same code gets automatically upgraded to the better memcpy when it's available. We know it works in practice and doesn't prevent ifunc selectors from doing their thing later. What's not to like? Sergey
