From Fedora Project Wiki
(Add priority=10 to mock configuration, to override system versions)
(Update list of instrumentations)
Line 28: Line 28:
** <s>Unsupported specifiers like `__declspec(thread)` can be parsed as old-style function definitions with a parameter implicitly typed as `int`, leading to a false positive implicit `int`.</s>  Should be fixed for `__declspec` in `.implicits.4.` builds.
** <s>Unsupported specifiers like `__declspec(thread)` can be parsed as old-style function definitions with a parameter implicitly typed as `int`, leading to a false positive implicit `int`.</s>  Should be fixed for `__declspec` in `.implicits.4.` builds.
** <s>A [https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107805 GCC bug] causes GCC to report an implicit `int` type if a type that has been redefined is used in a declaration.</s> Should be fixed in `.implicits.4.` builds.
** <s>A [https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107805 GCC bug] causes GCC to report an implicit `int` type if a type that has been redefined is used in a declaration.</s> Should be fixed in `.implicits.4.` builds.
* '''Implicit `int`/pointer conversion.''' Starting with the `.implicits.5.` versions, such errors are logged as well. The current implementation exhibits some false positives (even though cascading errors are suppressed in the `.implicits.6.` compiler), and there is currently no working suppression mechanism.
* '''Implicit `int`/pointer conversion.''' Starting with the `.implicits.5.` versions, such errors are logged as well. These are reported as '''int-conversion''' failures.
* '''Misuse of `return` statements.''' With the `.implicits.8.` versions, `return` without an expression in functions not returning `void` and vice versa are flagged as '''return-mismatch'''. Upstream, this is controlled by the new `-Wreturn-mismatch` warning, which was split from `-Wreturn-types`.
* '''Misuse of pointer types.''' With the `.implicits.8.` versions, implict casts by completely different pointer types (not involving the generic pointer-to-`void` types) are flagged as '''incompatible-pointer-types'''. Upstream, this is controlled by `-Wincompatible-pointer-types`.
* '''Misspelled type names in function prototypes.''' Reported as '''declaration-missing-parameter-type'''. Also in `.implicits.8.`. Previously, if all type names in a function parameter list were not recognized, the parameter list was ignored, and a warning (currently not controllable by a separate option) was issued.


The instrumentation in `redhat-rpm-config` checks after the various stages (e.g., `%build` and `%prep`) if `gcc` has logged anything.  If yes, it fails the build with a message like this one:
The instrumentation in `redhat-rpm-config` checks after the various stages (e.g., `%build` and `%prep`) if `gcc` has logged anything.  If yes, it fails the build with a message like this one:
Line 34: Line 37:
  <nowiki>
  <nowiki>
*** GCC errors begin ***
*** GCC errors begin ***
conftest.c:55: type defaults to int
conftest.c:55: implicit-int
conftest.c:37: implicit function declaration: exit
conftest.c:37: implicit-function-declaration: exit
conftest.c:83: defaulting to int
conftest.c:83: implicit-int
conftest.c:84: implicit function declaration: offsetof
conftest.c:84: implicit-function-declaration: offsetof
conftest.c:293: defaulting to int
conftest.c:293: implicit-int
conftest.c:71: type defaults to int
conftest.c:71: implicit-int
*** GCC errors end ***</nowiki>
*** GCC errors end ***</nowiki>


Line 70: Line 73:
A [https://copr.fedorainfracloud.org/coprs/fweimer/modernc-1/ COPR repository] contains instrumented packages. Builds fail if they use any of the obsolete features currently under test.
A [https://copr.fedorainfracloud.org/coprs/fweimer/modernc-1/ COPR repository] contains instrumented packages. Builds fail if they use any of the obsolete features currently under test.


{{warning}} From time to time, error checking will break if newer `gcc` or `redhat-rpm-config` builds arrive in rawhide.
{{warning}} From time to time, error checking will break if newer `gcc` or `redhat-rpm-config` builds arrive in rawhide, although the confuguration sets `priority=10` to prefer the instrumented builds.


The following `mock` configuration can be used for local builds.
The following `mock` configuration can be used for local builds.

Revision as of 11:26, 22 November 2023

Porting Fedora to Modern C

This page collects some technical details related to Changes/PortingToModernC.

Status

  • Gathering data on implicit function declarations (e.g., exit (0) called without #include <stdlib.h>).
  • Gathering data on implicit ints (e.g., const TWO = 2).
  • Preparing infrastructure for fix verification for Fedora developers.

Discovering Packages Which Need Fixing

Using Frédéric Bérat's Mass Prebuild tool, we build packages with instrumented gcc and redhat-rpm-config packages that fail the build if certain errors occur anywhere during the build. These forced failures are guaranteed to happen even if the build system ignores the exit status of a gcc command and continues regardless (for example, during configure checks). This is implemented by patching gcc in a special buildroot to log certain errors to a 1777 (sticky bit permission) directory. After the build, redhat-rpm-config checks if the directory is empty (or more precisely, only contains expected errors), and fails the build otherwise.

You can download logs from builds with this instrumented compiler by cloning a Git repository:

git clone --depth 1 https://gitlab.com/fweimer-rh/fedora-modernc-logs

Currently available instrumentations

  • Implicit function declarations. Limitations:
    • Some functions are expected to be undefined (e.g., pledge does not exist on Linux). Calls to them result in implicit declarations, which is a false positive in this context. As a result, in the redhat-rpm-config package, there is an exception list in the reporting code.
  • Implicit int (in variable declarations, and missing function return types). Limitations:
    • Unsupported specifiers like __declspec(thread) can be parsed as old-style function definitions with a parameter implicitly typed as int, leading to a false positive implicit int. Should be fixed for __declspec in .implicits.4. builds.
    • A GCC bug causes GCC to report an implicit int type if a type that has been redefined is used in a declaration. Should be fixed in .implicits.4. builds.
  • Implicit int/pointer conversion. Starting with the .implicits.5. versions, such errors are logged as well. These are reported as int-conversion failures.
  • Misuse of return statements. With the .implicits.8. versions, return without an expression in functions not returning void and vice versa are flagged as return-mismatch. Upstream, this is controlled by the new -Wreturn-mismatch warning, which was split from -Wreturn-types.
  • Misuse of pointer types. With the .implicits.8. versions, implict casts by completely different pointer types (not involving the generic pointer-to-void types) are flagged as incompatible-pointer-types. Upstream, this is controlled by -Wincompatible-pointer-types.
  • Misspelled type names in function prototypes. Reported as declaration-missing-parameter-type. Also in .implicits.8.. Previously, if all type names in a function parameter list were not recognized, the parameter list was ignored, and a warning (currently not controllable by a separate option) was issued.

The instrumentation in redhat-rpm-config checks after the various stages (e.g., %build and %prep) if gcc has logged anything. If yes, it fails the build with a message like this one:

*** GCC errors begin ***
conftest.c:55: implicit-int
conftest.c:37: implicit-function-declaration: exit
conftest.c:83: implicit-int
conftest.c:84: implicit-function-declaration: offsetof
conftest.c:293: implicit-int
conftest.c:71: implicit-int
*** GCC errors end ***

As a convenience, a dump of config.log files is included in the build log before these errors are listed.

Container image with preinstalled instrumented GCC

A container image with the GCC is available (source code). Run a shell with this command:

podman run -i -t registry.gitlab.com/fweimer-rh/fedora-modernc-container:modernc

By default, the instrumentation logs errors into the /usr/lib/gcc/errors directory.


Special Koji side tag of rawhide with instrumented GCC

If you have Fedora Koji build access, you can run a build with an instrumented GCC using this command:

  • koji build --scratch f39-toolchain source-RPM

Builds fail if they use any of the obsolete features currently under test.

The RPM buildroot repositories are available here:

They can be used in mock.

Special COPR repository with instrumented GCC

A COPR repository contains instrumented packages. Builds fail if they use any of the obsolete features currently under test.

From time to time, error checking will break if newer gcc or redhat-rpm-config builds arrive in rawhide, although the confuguration sets priority=10 to prefer the instrumented builds.

The following mock configuration can be used for local builds.

config_opts['target_arch'] = 'x86_64'
config_opts['legal_host_arches'] = ('x86_64',)

include('templates/fedora-rawhide.tpl')

config_opts['root'] = 'fedora-rawhide-modernc'

config_opts['dnf.conf'] += """

[copr-fweimer-modernc]
name=COPR with instrumented GCC for modern C porting
baseurl=https://download.copr.fedorainfracloud.org/results/fweimer/modernc-1/fedora-rawhide-$basearch/
type=rpm-md
skip_if_unavailable=False
gpgcheck=1
gpgkey=https://download.copr.fedorainfracloud.org/results/fweimer/modernc-1/pubkey.gpg
repo_gpgcheck=0
enabled=1
enabled_metadata=1
priority=10
"""

Fedora Package Updates

All identified issues need to be fixed in rawhide eventually. If a package cannot be ported, a bug needs to be filed in Bugzilla blocking PortingToModernCExemptions to record this, and the package still needs to build in an early C language mode (probably C89 or C99).

If an upstream fix becomes available, it may make sense to backport it before a package rebase, so that the package no longer shows up as needed porting worj.

Compatibility packages and other legacy packages need to be fixed as well.

Documenting Changes

  • Fedora dist-git commit message should refer Changes/PortingToModernC and Toolchain/PortingToModernC (this page).
  • Fedora updates should be captured in the fedora-modernc tracking repository. Please send MRs there if you fix something.
  • Changes that merely included what would come in through a future rebase do not need a bug in Bugzilla or filing an issue with upstream. This includes backports from upstream, or patches which merely change generated files when re-generation using the current tool version would also fix the issues. (The latter is common with old configure scripts.)
  • Patch files added as part of this change must include a URL for an upstream commit or for an upstream patch submission.
  • A bug needs to be filed in Bugzilla if there is no way to report the issue upstream, blocking PortingToModernCNoUpstream.
  • If a package cannot be ported, a bug needs to be filed blocking PortingToModernCExemptions.

Cross-Distribution Collaboration

A mailing list has been set up to exchange issues and observations of wider interest: c-std-porting@lists.linux.dev (archive)

Further resources will follow as needed.

Fedora Change Tracking

Changes to Fedora packages should be captured in the fedora-modernc tracking repository.

Not every build issue will have a bug in Bugzilla. But if there is no upstream with a mailing list or issue tracker to submit fixes to, it makes sense to create a bug in Bugzilla, especially for cross-distribution collaboration.

Currently, these trackers exist:

  • PortingToModernC Umbrella tracker for the tracker bugs below.
  • PortingToModernCExemptions Bugs which document that certain packages could not be switched to C99 or later language standards. Typically these packages keep building with -std=gnu89.
  • PortingToModernCHelpNeeded Bugs where help from the package maintainer is needed to address the issue(s).
  • PortingToModernCNoUpstream Bugs to provide a reference where there is either no upstream, or the current upstream lacks a public bug tracker or version control repository

Contributing To The Porting Effort

If you want to contribute to this effort, you should subscribe to the c-std-porting list mentioned above. If you create a Red Hat Bugzilla account, log in, and visit the Bugzilla tracker bugs linked above, a Add me to CC list check mark will appear on the bug pages. If you check that and click the Save Changes button, you will be notified of future bug changes, including bug dependency changes that indicate new issues. You can also subscribe to individual bugs that interest you.

At this point, it is too early to tell what level of knowledge and open-source experience is typically needed to successfully contribute to this project.