druntime: Test core.elf refactoring#2875
Conversation
|
Btw, from what I gathered, file I/O seems to be mandatory when looking up ELF sections. That's costly, so I gave up the original plan of replacing the magic linker symbols for the |
|
How costly can it be? If it's only done once when initializing druntime, it can't be that bad, can it? |
|
I haven't measured it, but opening the file for each D binary plus (quite) a bunch of mmap and munmap calls, all syscalls, and string comparisons (for each section, until we finally find the matching one) etc. doesn't seem very nice. |
|
What is it again that you need for Android TLS? The |
|
The |
|
Yeah, I thought of changing the section name in LLVM when responding now too, I'll see how bad it'd be to patch. |
|
|
|
Btw we could also get rid of the current LDC-specifics in |
|
Those symbols are indeed generic, but support varies across platforms – I'd advise looking at the linker source code if in doubt. |
|
Now upstream: dlang/druntime#2330 |
|
An updated version of this has landed in branch |
Also testing whether all CI services support testing submodule PRs from non-official repos. Edit: Okay, CircleCI doesn't support it, so nope, definitely not an option.