Weekly report 4, LLVM libc

Hello! This is a combined report for both week 3 and 4.

In these two weeks I’ve fixed several issues in LLVM libc, but quite a
lot of time has also been spent purely learning things. I will start
by going over what I’ve learned, and then refer to related issues.

To start with I have gotten quite comfortable with CVise, how to use
it and general tricks about writing the test script for determining
whether the issue is still there after reducing a source file. For
example, I had an issue about a the print format macro PRId64 not
being defined on LLVM libc

This caused an error that looked like this:

/home/cat/c/llvm-project/compiler-rt/lib/scudo/standalone/timing.h:182:22: error: expected ')'
Str.append("%14" PRId64 ".%" PRId64 "(ns) %-11s", Integral, Fraction, " ");

So my first attempt of reducing was to grep for “expected ‘)'”. This
went on to reduce the source file to simply: “(“. Maybe not the most
interesting thing, but it was the “aha-moment” for me with regards to
CVise, because what it did with the test script became clear.

To actually fix this issue I filed a bug
https://github.com/llvm/llvm-project/issues/63317 and got told by a
compiler-rt developer that the timing.cpp file is only used for
performance evaluation. So the temporary fix I made was to exclude it
from the build by checking for LLVM_LIBC_INCLUDE_SCUDO=ON in CMake
until the print format macros are added to LLVM libc.


The next thing I’ve learned a lot about is C++ and standard C header
interoperability, or “include hell”. I learned about the differences
between C++ standard headers like “cwhatever” and “whatever.h”, also
what #include_next did, and also that compilers ship their own header
files like stddef.h and inttypes.h.

I first ran into this when pulling new commits from master and rebuilt
LLVM libc, thinking that the errors were related to this. Weirdly
enough the original error just went away, and I couldn’t reproduce it
at all. But I quickly ran in to a similar issue when compiling LLVM
libc in fullbuild mode on a llvm/musl system.

This time it was an error about wint_t not being defined:

> /llvm-project/build-libc-full/projects/libc/include/wchar.h:21:11: error: unknown type name 'wint_t'
> int wctob(wint_t) __NOEXCEPT;

The issue here arrises because LLVM libc’s llvm-libc-types/wint_t.h
gets the wint_t type using:

> #define __need_wint_t
> #include
> #undef __need_wint_t

This depends on internal behaviour of the stddef.h header. Because
this is C++ it will include in this case libc++’s stddef.h, but this
#include_next’s the second stddef.h in the include search path.

glibc uses __need_wint_t to make stddef.h define wint_t, while musl
uses __NEED_wint_t. No one is wrong here, as it is libc internals that
should not be used by end users, instead something like wchar.h should
be included. However, as this is a libc implementation too it does not
make sense to include all of that stuff, so something else must be
done. I then grepped the whole llvm checkout for stddef.h and realized
that Clang shipped its own stddef.h too. This header, like glibc, uses
__need_wint_t to define wint_t, which is exactly what I want. I posted
a bug report and got told that the internal Clang headers are to be
used, not the system libc’s headers, because of issues like these.

However, somehow /usr/include is higher up in the include order than
/usr/lib/clang/* even when using -ffreestanding, so I assume this is a
bug with the Gentoo llvm/musl stage actually.

Another thing I have worked on is to replace some __unix__ ifdef
checks in LLVM libc with __linux__. When looking through the source
code there are quite a lot of places where these are mixed up
bizarrely enough. The most obvious ones are __unix__ check followed by
an #include . This is caused by “cargo cult” meaning that
they once did it that way and stuck with it for no reason.

I have fixed this here: https://reviews.llvm.org/D153729#4447435, but
I will revisit this because there’s a chance that macOS users could
have used the typedefs pthread_once_t and once_flag, even though the
underlying type __futex_word was supposed to be Linux only, because
futexes here are Linux kernel specific. This would’ve
previously not have errored out on macOS (__APPLE__) since __unix__ is
defined, and __futex_word is just defined as an aligned 32 bit uint,
no unconditional kernel headers used that would’ve broken the build.

I will therefore go back and define these for macOS later.

I have also done some work on upstreaming things needed for Python
into LLVM libc instead of just mashing everything into my Python
source dir. The first one being the POSIX extension fdopen().

As fopen is already implemented the hard part was not the function
itself, but actually figuring out where everything in LLVM libc was
placed. Apart from the obvious declaration in the internal headers,
and corresponding source file, I also needed to make sure it was
usable in the libc. In total I needed to edit 7 files, like the
TableGen specifications, config/$arch/entrypoints.txt, libc
exposed internal header file, and of course CMakeLists.txt.

This is not upstreamed yet but I am working on it here

The other thing I want to get upstreamed is the limits.h header, in my
case needed for SSIZE_MAX. I have successfully made a tiny version in
my libc tree that exposes some macros, and I will try to upstream what
I have and then work on things one by one. Similarly here, the hard
part was actually getting the header and macros to be exposed in the
libc by editing build system code and specification files. I could
have temporarily just jammed in a limits.h header file but I think
it’s important to get to know how LLVM libc does things/”how the
boilerplate works” early in my project.

That’s all the big things, I also continued work on Python, fixed some
small stuff like a typo fix
adding Emacs support in llvm-common
and other “Gentoo but not really GSoC work”:
https://github.com/gentoo/gentoo/pull/31560 (license fix, soju).
https://github.com/gentoo/gentoo/pull/30933 (new package, senpai).

Next week I will work on getting Clang/LLVM supported in
Crossdev. This will be done by first making sure that the hosts LLVM
toolchain supports the target architecture via the LLVM_TARGETS
USE-flag. Currently Clang on Gentoo can compile things like the
kernel, but anything that relies on runtime libraries, like libc,
fails due to compiler-rt not being compiled for the target triple, so
I will also make sure that Crossdev compiles compiler-rt for the
specified target triple.

– —

This entry was posted in Bootstapping LLVM. Bookmark the permalink.

Leave a Reply

Your email address will not be published.