Qt5-base 5.15 build fails with llvm+clang

dilyn

I'm attempting to build qt5-base on a Linux machine. I previously could build successfully using GCC 10. I recently switched the whole system to LLVM (clang, lld, etc.) and am hitting a build error I'm not sure how to tackle.
Below is the build log.
http://ix.io/2oUD

The specific error occurs in QtNetwork:

src/network/socket/qnet_unix_p.h:120:29: error: cannot initialize a parameter of type 'socklen_t *' (aka 'unsigned int *') with an rvalue of type 'int *'

Unfortunately I don't know C(PP), do I can't particularly parse this.

I'm using llvm 10. To avoid other errors I found I had to manually specify the platform as linux-clang-libc++.

SGaist

Hi and welcome to devnet,

Can you give more information about your Linux setup ?
What distribution ?
What package did you install ?
What configure options did you use exactly ?

dilyn

I'm using KISS linux, just a simple musl-based, source distro.
I'm building qt5 from source (qt5-base-everywhere-src-5.15.0.tar.xz); this is exactly what I'm doing to build it:

# Don't link against execinfo.h.
sed -i 's/define QLOG/define N/g' \
    ./src/corelib/global/qlogging.cpp

./configure \
    -confirm-license \
    -opensource \
    -prefix /usr \
    -docdir /usr/share/doc/qt \
    -headerdir /usr/include/qt \
    -archdatadir /usr/lib/qt \
    -datadir /usr/share/qt \
    -sysconfdir /etc/xdg \
    -examplesdir /usr/share/doc/qt/examples \
    -nomake examples \
    -nomake tests \
    -platform linux-clang-libc++ \
    -optimized-qmake \
    -no-libudev \
    -no-separate-debug-info \
    -no-pch \
    -no-dbus \
    -no-accessibility \
    -no-gtk \
    -no-glib \
    -system-libjpeg \
    -system-libpng \
    -system-sqlite \
    -system-zlib \
    -system-freetype \
    -system-harfbuzz \
    -silent

make
make INSTALL_ROOT="$1" install

These are the dependencies we have listed for qt5-base (although from testing it seems ICU is also required; but that eror message occurs in Widgets)

bison make
flex  make
freetype-harfbuzz
gperf make
libICE
libSM
libX11
libXext
libXi
libXrandr
libXrender
libXslt
libinput
libjpeg-turbo
libpng
libxkbcommon
linux-headers make
mesa
mtdev
perl  make
sqlite
xcb-util
xcb-util-image
xcb-util-keysyms
xcb-util-wm
xcb-util-renderutil
zlib

The only real thing that has changed on my system since the last time I built qt5 is dropping gcc for llvm (clang, libc++, compiler-rt, lld, etc). I've rebuilt basically the entire system after this with only a few hiccups.

I've tried choosing different platforms as well. linux-clang results in an identical error message. Interestingly, autodetection acts as if I'm using gcc still. Which may be a clue?

We DO use libressl instead of openssl (and indeed we apply patches to qt5 to enable ssl support), but considering I've built 5.15 previously with gcc that doesn't appear to be the issue.

SGaist

I would check the bug report system. They might have not used LLVM 10.

Did you check with LLVM 9 ?

dilyn

Thanks for the link; I have posted a bug report.

I have not tested with llvm 9; I will try it out, it'll take a while to see though. I'll report back when it's done.

dilyn

Update: also fails with an identical error using llvm 9.0.1.

I'm going to continue troubleshooting this, see if I can get some other people to reproduce it.

WMH1

I have just experienced what looks like the exact same error. Even the same line number. Clang++-19, Qt5-base 5.15.17, Linux Alpine 3.22

../../include/QtNetwork/5.15.17/QtNetwork/private/../../../../../../src/network/socket/qnet_unix_p.h:120:29: error: cannot initialize a parameter of type 'socklen_t *' (aka 'unsigned int *') with an rvalue of type 'int *'
  120 |     fd = ::accept4(s, addr, static_cast<QT_SOCKLEN_T *>(addrlen), sockflags);
      |                             ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

I believe it is a result of the defines in qtplatformdefs.h, where linux clang & linux llvm are implemented with opposite logic to linux g++? (Alpine will be the default #else in both cases, because it is MUSL, not GLIBC. One will be socklen_t and the other will be the offending int. Swapping the logic to match g++ is probably a hack, but it then compiles ok on Alpine.)