Safe signal handling around system calls is mandatory for user-mode
emulation, and requires a small piece of handwritten assembly code.
So refuse to compile unless the common-user/host subdirectory exists
for the host architecture that was detected or selected with --cpu.

Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

It is already included in the symlink shell function.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

container_hosts is matched against $cpu, so it must contain QEMU
canonical architecture names, not Debian architecture names.
Also do not set $container_hosts inside the loop, since it is
already set before.

Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Using the new ensuregroup command, the desired versions of meson and
sphinx can be placed in pythondeps.toml rather than configure.

The meson.install entry in pythondeps.toml matches the version that is
found in python/wheels.  This ensures that mkvenv.py uses the bundled
wheel even if PyPI is enabled; thus not introducing warnings or errors
from versions that are more recent than the one used in CI.

The sphinx entries match what is shipped in Fedora 38.  It's the
last release that has support for older versions of Python (sphinx 6.0
requires Python 3.8) and especially docutils (of which sphinx 6.0 requires
version 0.18).  This is important because Ubuntu 20.04 has docutils 0.14
and Debian 11 has docutils 0.16.

"mkvenv.py ensure" is only used to bootstrap tomli.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Debian only introduced tomli in the bookworm release.  Use a
vendored wheel to avoid requiring a package that is only in
bullseye-backports and is also absent in Ubuntu 20.04.

While at it, fix an issue in the vendor.py scripts which does
not add a newline after each package and hash.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Since there is a vendored copy, there is no point in choosing online
operation.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

The tricore tools are not detected when they are installed in
the host system, only if they are taken from an external
container.  For this reason the build-tricore-softmmu job
was not running the TCG tests.

In addition the container provides all tools, not just as/ld/gcc,
so there is no need to special case tricore.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

The CPU model has to be canonicalized to what Meson wants in the cross
file, to what Linux uses for its asm-$ARCH directories, and to what
QEMU uses for its user-mode emulation host/$ARCH directories.  Do
all three in a single case statement, and check that the Linux and
QEMU directories actually exist.

At a small cost in repeated lines, this ensures that there are no hidden
ordering requirements between the case statements.  In particular, commit
89e5b793 ("configure: Fix linux-user host detection for riscv64",
2023-08-06) broke ppc64le because it assigned host_arch based on a
non-canonicalized version of $cpu.

Reported-by: Joel Stanley <joel@jms.id.au>
Fixes: 89e5b793 ("configure: Fix linux-user host detection for riscv64", 2023-08-06)
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Acked-by: Ilya Leoshkevich <iii@linux.ibm.com>
Reviewed-by: Michael Tokarev <mjt@tls.msk.ru>
Tested-by: Ilya Leoshkevich <iii@linux.ibm.com>
Tested-by: Michael Tokarev <mjt@tls.msk.ru>
Message-ID: <20230808120303.585509-4-pbonzini@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

x32 uses the same signal handling fragments as x86_64, since host_arch
is set to x86_64 when Meson runs.  Remove the unnecessary forwarder and
set the host_arch variable properly in configure.

Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Acked-by: Ilya Leoshkevich <iii@linux.ibm.com>
Reviewed-by: Michael Tokarev <mjt@tls.msk.ru>
Tested-by: Ilya Leoshkevich <iii@linux.ibm.com>
Tested-by: Michael Tokarev <mjt@tls.msk.ru>
Message-ID: <20230808120303.585509-2-pbonzini@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Mirror the host_arch variable from meson.build, so that we
probe for the correct linux-user/include/host/ directory.

Fixes: e3e477c3 ("configure: Fix cross-building for RISCV host")
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

This can be useful for setting some meson global options, such as the
optimization level or debug state.xs

Signed-off-by: Daniel P. Berrangé <berrange@redhat.com>
Message-Id: <20230801130403.164060-7-berrange@redhat.com>
[thuth: Move the help text into the section with the other --... options]
Signed-off-by: Thomas Huth <thuth@redhat.com>

Update $linux_arch to keep using the shared linux-headers/asm-riscv/
include path.

Fixes: e3e477c3 ("configure: Fix cross-building for RISCV host")
Reported-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org>
[rth: Missed v5, so now applying the diff between v4 and v5.]
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

While when building on native Linux the host architecture
is reported as "riscv32" or "riscv64":

  Host machine cpu family: riscv64
  Host machine cpu: riscv64
  Found pkg-config: /usr/bin/pkg-config (0.29.2)

Since commit ba0e7333 ("configure: Merge riscv32 and riscv64
host architectures"), when cross-compiling it is detected as
"riscv". Meson handles the cross-detection but displays a warning:

  WARNING: Unknown CPU family riscv, please report this at https://github.com/mesonbuild/meson/issues/new
  Host machine cpu family: riscv
  Host machine cpu: riscv
  Target machine cpu family: riscv
  Target machine cpu: riscv
  Found pkg-config: /usr/bin/riscv64-linux-gnu-pkg-config (1.8.1)

Now since commit 278c1bce ("target/riscv: Only unify 'riscv32/64'
-> 'riscv' for host cpu in meson") Meson expects the cpu to be in
[riscv32, riscv64]. So when cross-building (for example on our
cross-riscv64-system Gitlab-CI job) we get:

  WARNING: Unknown CPU family riscv, please report this at https://github.com/mesonbuild/meson/issues/new


  Host machine cpu family: riscv
  Host machine cpu: riscv
  Target machine cpu family: riscv
  Target machine cpu: riscv
  ../meson.build:684:6: ERROR: Problem encountered: Unsupported CPU riscv, try --enable-tcg-interpreter

Fix by partially revert commit ba0e7333 so when cross-building
the ./configure script passes the proper host architecture to meson.

Fixes: ba0e7333 ("configure: Merge riscv32 and riscv64 host architectures")
Fixes: 278c1bce ("target/riscv: Only unify 'riscv32/64' -> 'riscv' for host cpu in meson")
Reported-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Message-Id: <20230711110619.56588-1-philmd@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

The call to git-submodule.sh done in configure may happen without a
previous checkout of the roms/SLOF submodule, or even without a
previous run of the script.

So, handle creating a .git-submodule-status file even in validate
mode.  If git is absent, ensure that all passed directories exists
(because you should be in a fresh untar and will not have stale
arguments to git-submodule.sh) but do no other checks.  If git
is present, ensure that .git-submodule-status contains an entry
for all submodules passed on the command line.

With this change, "ignore" mode is not needed anymore.

Reported-by: Nina Schoetterl-Glausch <nsg@linux.ibm.com>
Fixes: b11f9bd9 ("configure: move SLOF submodule handling to pc-bios/s390-ccw", 2023-06-06)
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

In commit b0fcc6fc ("build: rebuild build.ninja using
"meson setup --reconfigure"", 2023-05-19) I changed the build.ninja
rule in the Makefile to use "meson setup" so that the Makefile would
pick up a changed path to the meson binary.

However, there was a reason why build.ninja was rebuilt using $(NINJA)
itself.  Namely, ninja has its own cache of file modification times,
and if it does not know about the modification that was done outside
its control, it will *also* try to regenerate build.ninja.  This can be
simply by running "make" on a fresh tree immediately after "configure";
that will trigger an unnecessary meson run.

So, apply a refinement to the rule in order to cover both cases:

- track the meson binary that was used (and that is embedded in
  build.ninja's reconfigure rules); to do this, write build.ninja.stamp
  right after executing meson successfully

- if it changed, force usage of "$(MESON) setup --reconfigure" to
  update the path in the reconfigure rule

- if it didn't change, use "$(NINJA) build.ninja" just like before
  commit b0fcc6fc.

Reported-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Tested-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

this allows us to exercise the startup code used by GCC to call main().

Signed-off-by: Bastian Koppelmann <kbastian@mail.uni-paderborn.de>
Message-Id: <20230526061946.54514-4-kbastian@mail.uni-paderborn.de>

If configure was invoked with --disable-download and git
submodules were not checked out a warning is produced and the
configure script fails. But the $download variable (which
reflects the enable/disable download argument) is checked for in
a weird fashion:

  test -f "$download" = disabled

Drop the '-f' to check for the actual value of the variable.

Fixes: 2019cabf ("meson: subprojects: replace submodules with wrap files", 2023-06-06)
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Reuse --enable/--disable-download to control git submodules as well.
Adjust the error messages of git-submodule.sh to refer to the new
option.

Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

The only remaining user of submodules at build time is roms/SLOF,
which is handled in pc-bios/s390-ccw/Makefile.  Remove the relevant
code from the main makefile.

Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Unlike other subprojects, these require an overlay directory to include
meson rules to build the libraries.  The rules are basically lifted
from tests/fp/meson.build, with a few changes to create platform.h
and publish a dependency.

The build defines are passed through a subproject option, and posted
back to users of the library via the dependency's compile_args.

The only remaining user of GIT_SUBMODULES and GIT_SUBMODULES_ACTION
is roms/SLOF, which is used to build pc-bios/s390-ccw.  All other
roms submodules are only present to satisfy the license on pre-built
firmware blobs.

Best reviewed with --color-moved.

Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Move the handling of the roms/SLOF submodule out of the main Makefile,
since we are going to remove submodules from the build process of QEMU.

Acked-by: Thomas Huth <thuth@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Compared to submodules, .wrap files have several advantages:

* option parsing and downloading is delegated to meson

* the commit is stored in a text file instead of a magic entry in the
  git tree object

* we could stop shipping external dependencies that are only used as a
  fallback, but not break compilation on platforms that lack them.
  For example it may make sense to download dtc at build time, controlled
  by --enable-download, even when building from a tarball.  Right now,
  this patch does the opposite: make-release treats dtc like libvfio-user
  (which is not stable API and therefore hasn't found its way into any
  distros) and keycodemap (which is a copylib, for better or worse).

dependency() can fall back to a wrap automatically.  However, this
is only possible for libraries that come with a .pc file, and this
is not very common for libfdt even though the upstream project in
principle provides it; it also removes the control that we provide with
--enable-fdt={system,internal}.  Therefore, the logic to pick system
vs. internal libfdt is left untouched.

--enable-fdt=git is removed; it was already a synonym for
--enable-fdt=internal.

Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

The behavior of --{enable,disable}-pypi is similar to that of
-Dwrapmode={default,nodownload} respectively.  In particular,
in both cases a feature needs to be explicitly enabled for the
dependency to be downloaded.

So, use a single option to control both cases.  Now, --enable-slirp
will trigger cloning and building of libslirp if the .pc file
is not found on the machine.

Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Reviewed-by: Thomas Huth <thuth@redhat.com>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

The scenario for which --with-git= was introduced was to use a SOCKS proxy
such as tsocks.  However, this was back in 2017 when QEMU's submodules
used the git:// protocol, and it is not as important when using the
"smart HTTP" backend; for example, neither "meson subprojects download"
nor scripts/checkpatch.pl obey the GIT environment variable.

So remove the knob, but test for the presence of git in the configure and
git-submodule.sh scripts, and suggest using --with-git-submodules=validate
+ a manual invocation of git-submodule.sh when git does not work.  Hopefully
in the future the GIT environment variable will be supported by Meson.

Reviewed-by: Thomas Huth <thuth@redhat.com>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Setting the MAKE variable to a GNU Make executable does not really have
any effect: if a non-GNU Make is used, the QEMU Makefile will fail to
parse.  Just remove everything related to --make and $make as dead code.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

By using a subproject, our own meson.build can use variables from
the subproject instead of hard-coded paths.  This is also the first step
towards managing downloads with .wrap files instead of submodule.

Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Recent dtc/libfdt can use either Make or meson as the build system.
By using a subproject, our own meson.build can remove the hard
coded list of source files.

This is also the first step towards managing downloads with .wrap
files instead of submodule.

Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Apart from CLICOLOR_FORCE and GREP_OPTIONS, there are other variables
that are listed in the Autoconf manual.  While Autoconf neutralizes them
very early, and assumes it does not (yet) run in a shell that has "unset",
QEMU assumes that the user invoked configure under a POSIX shell, and
therefore can simply use "unset" to clear them.

CDPATH is particularly nasty because it messes up "cd ... && pwd".

Reported-by: Juan Quintela <quintela@redhat.com>
Cc: qemu-stable@nongnu.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Some scripts are invoked via the first "python3" binary in the PATH,
because they are executable and their shebang line is "#! /usr/bin/env
python3".  To enforce usage of $(PYTHON), make them nonexecutable.
Scripts invoked via meson need nothing else, and meson-buildoptions.py
is already using $(PYTHON).  For probe-gdb-support.py however the
invocation in the configure script has to be adjusted.

Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Reintroduce the cmd_line.txt mangling to remove the sphinx_build option
when rerunning meson.  The mechanism was removed in commit 75cc2864
("configure: remove backwards-compatibility code", 2023-01-11) because
the fixups were obsolete at the time; however, the Meson deprecation
mechanism doesn't quite work when options are finally removed, so we
need to bring it back.

Reported-by: Peter Maydell <peter.maydell@linaro.org>
Tested-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Add support for the ELF flags
Move target/hexagon/cpu.[ch] to be v73
Change the compiler flag used by "make check-tcg"

The decbin instruction is removed in Hexagon v73, so check the
version before trying to compile the instruction.

Signed-off-by: Taylor Simpson <tsimpson@quicinc.com>
Reviewed-by: Anton Johansson <anjo@rev.ng>
Message-Id: <20230427224057.3766963-2-tsimpson@quicinc.com>

All calls to probe_target_compiler are conditioned on
some "have_target" invocation, or inside a loop on target_list.
Therefore there is no issue with building unnecessary
firmware images and tests.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Move some variable assignments around for clarity and to remove
one of three loops on the command line arguments.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

It is taken care of by the symlink shell function.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Tests run in configure are pretty trivial at this point, so
do not bother with the extra complication of running tests
both with and without -Werror.

Reviewed-by: Marc-André Lureau <marcandre.lureau@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

The comment is not correct anymore, in that the usability test for
the compiler and linker are done after probing $cpu, and Meson will
redo them anyway.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Use the slighly nicer .version_compare() function for GCC; for Clang that is
not possible due to the mess that Apple does with version numbers.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Remove the only remaining uses of QEMU_CFLAGS.  Now that no
feature tests are done in configure, it is possible to remove
CONFIGURE_CFLAGS and CONFIGURE_LDFLAGS as well.

Reviewed-by: Marc-André Lureau <marcandre.lureau@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Meson already knows to test with the positive form of the flag, which
simplifies the test.  Warnings are now tested explicitly for the C++
compiler, instead of hardcoding those that are only available for
the C language.

At this point all compiler flags in QEMU_CFLAGS are global and only
depend on the OS.  No feature tests are performed in configure.

Reviewed-by: Marc-André Lureau <marcandre.lureau@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>