Features/Meson: Difference between revisions
(→Easy) |
|||
Line 49: | Line 49: | ||
:Move the build rules from tests/Makefile.include to tests/meson.build. You can take inspiration from tests/qtest/Makefile.include. | :Move the build rules from tests/Makefile.include to tests/meson.build. You can take inspiration from tests/qtest/Makefile.include. | ||
:[https://github.com/bonzini/qemu/tree/meson-poc Here] you can also find patches to convert the build rules in roms/ and pc-bios/ | :[https://github.com/bonzini/qemu/tree/meson-poc Here] you can also find patches to convert the build rules in roms/ and pc-bios/ | ||
:[https://github.com/elmarco/qemu/blob/meson-poc/tests/meson.build Here] you can find some old/poc conversion of tests/ | |||
:'''Benefit:''' allows removing CFLAGS, LDFLAGS, LIBS and LIBS_SOFTMMU, as well as most if not all of rules.mak. | :'''Benefit:''' allows removing CFLAGS, LDFLAGS, LIBS and LIBS_SOFTMMU, as well as most if not all of rules.mak. | ||
;Relocatable install | ;Relocatable install |
Revision as of 16:29, 22 August 2020
For the rationale behind the conversion and the development notes, see Features/Meson/Design.
Summary of changes
- New system requirements
- GNU Make 3.82
- Python setuptools
- Meson 0.55.0 (in order to use the system-wide install).
- Source code changes
- You cannot #include "config-target.h" and "config-devices.h". Instead, you #include CONFIG_TARGET and CONFIG_DEVICES (i.e. filenames are specified via macros; meson.build takes care of defining the macros).
- .c files (including .inc.c files) cannot be #included. They have been renamed to .c.inc, for consistency with existing .rst.inc files.
- The name of files produced by decodetree is fixed (the C code for XYZ.decode ends up in decode-XYZ.c.inc).
- Subdirectories with trace events require a forwarding header file; see hw/scsi/trace.h for an example.
- Build layout changes
- The Makefile is not recursive anymore. You don't make aarch64-softmmu/all or make arm-linux-user/all, you make qemu-system-aarch64 or make qemu-arm.
- Emulator binaries have moved to the root of the build directory. For now, paths such as cris-softmmu/qemu-system-cris are preserved for backwards compatibility through symlinks (created by configure). They are just a user convenience however. You cannot "rm" them to force a rebuild, for example.
- Other binaries have moved to the directory where they reside in the source tree. For example, "virtiofsd" has moved to "tools/virtiofsd/virtiofsd".
- Most binaries are built by default by make all, including several in contrib/ that weren't built before. Currently, vhost-user-blk and rdmacm-mux are not because they fail on 32-bit and big-endian platforms respectively.
- Sphinx manuals are only built if their constituent files are changed, therefore "make sphinxdocs" will be plenty fast when you've only modified a file or two.
- Dependencies added to libqemuutil.a will propagate to all programs that link to it. If util/ code has dependencies, it should be conditional on have_system or have_block if applicable (this was a bug in the Makefiles).
- For bisection: incremental builds across the conversion work fine in the forwards direction. They probably don't work at all backwards across the conversion, so you'll need to throw away your build tree when bisection moves you backwards across the conversion. This is mitigated by the fact that bisection usually starts at a release, and the patches have been applied on top of the 5.1.0 tag.
- New conventions
- New configure options should also be added to meson_options.txt so that dependencies are tested in meson.build rather than configure. How to do so is documented in docs/devel.
- Impact to developers
- If you were already using out-of-tree (VPATH) builds, nothing changes; an incremental pull of these changes should still build.
- If you are used to in-tree builds, you'll want to do make distclean prior to merging in these changes (if you forget, git will remind you about various trace.h files that still exist as built files in your in-tree build vs. checked-in files post-patch). After that one-time clean, you can then proceed to do ./configure && make as before, but things will now automatically create a subdirectory build/ on your behalf where the actual build is performed.
CI holes
The following issues were _not_ found by GitLab CI:
- SDL 2.0.8 requires -Wno-undef
- s390x-softmmu is special for cross-compilation, because it is the only target to use the host C compiler.
- There is special code in configure to handle a C++ compiler that does not play well with the C compiler.
- Not all linux-user TCG tests are covered.
- Big endian platforms are not covered.
- ARM host platforms are not covered.
- 32-bit POSIX platforms are not covered.
- --disable-tools --enable-system builds are not covered.
- Peter uses Ubuntu+MXE instead of Fedora-mingw to build cross compilers. docker-test-mingw only works for Fedora and should be made more generic. Also container images for Fedora-mingw should be prepared in the same way as the existing container images for Debian+MXE.
Blockers for further conversions
- Complete moving installation rules to meson.build
- See here. You're welcome to propose yourself for helping!
- Benefit: allows removing parts of Makefile, leaving it as just an orchestrator for submodules, Makefile.ninja and Makefile.mtest
- Complete moving build rules to meson.build
- Move the build rules from tests/Makefile.include to tests/meson.build. You can take inspiration from tests/qtest/Makefile.include.
- Here you can also find patches to convert the build rules in roms/ and pc-bios/
- Here you can find some old/poc conversion of tests/
- Benefit: allows removing CFLAGS, LDFLAGS, LIBS and LIBS_SOFTMMU, as well as most if not all of rules.mak.
- Relocatable install
- Not really related to Meson, but it allows removing the pre_prefix hack in the configure script.
- Benefit: on Windows only /usr/share/qemu is relocatable, everything else (icons etc.) is not. POSIX is currently not relocatable and would get it for free.
Easy
- Diagnose "make is too old" in configure (or in the makefile?)
- In particular OSX system make is too old, so we'll get a lot of confused users there otherwise, because the error message is pretty obscure ("multiple target patterns"). Possibly checking in the makefile itself would be nicer, to catch the situation where the user forgets and runs 'make' when they meant 'gmake'.
- Move pkg-config tests to meson.build
- Replace declare_dependency statements with dependency statements
- Benefit: allows removing *_CFLAGS and *_LIBS variables from config-host.mak. Allows using dependency objects instead of CONFIG_* symbols when declaring sourcesets.
- Change simple config-host.mak symbols to Meson options
- Some symbols (e.g. CONFIG_PARALLELS) are only used placed in config-host.mak to be read from Meson. Instead pass them through -D... command-line options
- Benefit: allows removing CONFIG_* variables from config-host.mak.
Medium
- "Handle (faked-up) in-tree build
- And cleanup code to handle in-tree builds in configure. Or fail gracefully and explain the situation (which would be easier).
- Benefit: simpler upgrade path
- Include config-target.mak files in the tree
- Remove the configuration loop in configure and just include the content of the .mak files directly in the tree, for example in a target/configs/ directory. (There are some small complications because config-target.mak file currently includes the host disassembler symbols too, e.g. CONFIG_I386_DIS).
- Benefit: remove large swaths of imperative code, remove *-linux-user and *-softmmu directories from the build tree.
- Move other compile tests from configure to meson.build
Hard(er)
- Create configure help and command line parsing code from meson_options.txt.
- Write a program that takes the output of "meson introspect --buildoptions" and generates shell code to initialize variables, parse command line options, print the help and create -D command line options for meson. May only make sense once many or most options are converted.
- Benefit: remove code duplication, centralize command line handling
TBD
- Get rid of ninjatool and just require Ninja
- Benefit: 1000 lines of code go away, can use things such as ninja -t clean <TARGET>
- Disadvantage: recursive build re-enters through the window