Changes/LTO by default for package builds
Summary
This is a proposal to enable link time optimization (LTO) of packages built with rpmbuild by default. It's an over-simplification, but think of LTO as deferring analysis, optimization and code generation until creation of an executable or dynamic shared object.
This is implemented by adding the option "-flto" the injected flags in redhat-rpm-config. There will be a simple way for packages to opt-out of LTO.
Owner
- Name: Jeff Law
- Email: law@redhat.com
Current status
- Targeted release: Fedora 33
- Last updated: 2020-04-21
- Tracker bug: #1789115
- Release notes tracker: #429
Detailed Description
Programs built with rpmbuild and which honor flags injection via redhat-rpm-config will be built with LTO by default. A simple opt-out mechanism will be provided for packages which use features that are not LTO compatible.
The LTO bytecode itself will not be distributed as it is not stable from one GCC release to the next. This is enforced by stripping the LTO bytecode from any installed .o/.a files. We'll use bits SuSE has already written for redhat-rpm-config to implement this. (RFE filed for verifying this in rpminspect)
Minor changes are desirable to the %configure macro in redhat-rpm-config to fix common code idioms used by autoconf generated scripts which are compromised by the additional optimization enabled by LTO. Minor updates to various packages will be needed to opt-out of LTO or fix bugs exposed by LTO.
Bugs for the 3 issues we should address in redhat-rpm-config are here:
https://bugzilla.redhat.com/show_bug.cgi?id=1789099 https://bugzilla.redhat.com/show_bug.cgi?id=1789137 https://bugzilla.redhat.com/show_bug.cgi?id=1789149
Benefit to Fedora
The primary benefits of building with LTO enabled are smaller, faster executables/DSOs. A secondary benefit is LTO allows deeper analysis of package source code at compile time which can improve various GCC diagnostics and thus improve our ability to catch bugs at compile time such as uninitialized objects, buffer overflows, unterminated strings, restrict violations, etc.
This change also brings us back on-par with openSUSE who enabled LTO by default for their free distribution earlier in 2019.
If you're interested in some of the performance data and lower level details:
http://www.ucw.cz/~hubicka/slides/opensuse2018-e.pdf
https://hubicka.blogspot.com/2019/05/gcc-9-link-time-and-inter-procedural.html
And opensuse's bug tracker for their LTO enablement work:
https://bugzilla.opensuse.org/1133084
Scope
- Proposal owners:
The primary change is to redhat-rpm-config to add LTO to the default compile/link flags as well as a conditional which allows easy opt-out on a package by package basis. Additionally the post-build scripts need to strip the LTO bytecodes from any installed .o/.a files.
Additionally, we know there are many packages with configure scripts that are compromised by LTO. I have tweaks to the %configure macro in redhat-rpm-config which fixes the vast majority of these problems with a few simple sed scripts on the generated output. Like the basic support for injecting the LTO flags, this will require coordination with the redhat-rpm-config maintainers. Packages which call configure directly and have compromised tests will need a one line change to their .spec files to fix their configure scripts.
Some packages will need to opt-out of using LTO at this time. The most common case are packages that use symbol versioning or toplevel ASM statements. While there is a new mechanism to make LTO work with symbol versioning, I don't think any packages have been updated to use that mechanism. This will require a one line change to 50-75 packages (my script to find these is still running).
Finally, some packages will fail to build with LTO due to deeper analysis for compile-time diagnostics catching programming mistakes that have gone unnoticed until now. I'll obviously be working with package maintainers on all of these issues.
Note that even though the changes are fairly well localized in redhat-rpm-config and a small number of packages, the real scope of this change is much larger since it affects all packages in the distribution that are compiled with GCC and which honor the flags injection by redhat-rpm-config.
- Other developers:
As I mentioned, I'm happy to contact package owners that need to modify their packages and suggest how their package needs to be fixed. As a multi-decade GCC developer, I'm particularly well suited to describe LTO, its limitations and how LTO impacts the diagnostics from GCC to any package owner that needs additional information.
I'm also capable and available to address any GCC issues that we may arise as a result of this change. I don't expect much of the latter as SuSE has already enabled this feature for their distribution and thus weeded out most of the issues.
The highest level of coordination will be with the redhat-rpm-config maintainers.
I will also be coordinating with the GDB team to address debugging issues related to LTO. The most important issue is to ensure that we can pass the GDB testsuite with and without the -flto option being enabled. Failure to meet this goal would be considered a blocking issue for LTO enablement.
I'm also already in contact with SuSE and Debian/Ununtu engineers to discuss issues with gcc-10 with and without LTO.
We know there are some problems with debugging LTO code. I will be working with the GDB team to identify these issues and fix them either in the debugger or compiler as needed.
I have prototype code for the required redhat-rpm-config changes and I'll coordinate with the redhat-rpm-config maintainer to get them into the desired final form.
I also know every package that fails with LTO enabled. I'm still categorizing those failures. Many will ultimately need to use the opt-out mechanism because they use features that are not compatible with LTO. I expect to have all this ready to go the first work week of the new year. I will coordinate with package owners to either add the opt-out markers or fix issues in the package as needed.
- Release engineering: [ https://pagure.io/releng/issue/9119 ]
Aside from the redhat-rpm-config changes to ensure we do not ship LTO bytecode in .o/.a files, I do not expect any work from releng to be necessary. However, they need to be aware of the change and who to contact in case of issues. The redhat-rpm-config changes necessary to implement stripping of LTO sections/symbols can be found attached to this BZ:
https://bugzilla.redhat.com/show_bug.cgi?id=1789099
We will not do a special mass rebuild for this feature. However a mass rebuild is typically planned for the introduction of a new compiler. We would want to piggy back on that mass rebuild.
- Policies and guidelines:
It would be useful to document how to opt-out of LTO in the packaging guidelines.
- Trademark approval: N/A (not needed for this Change)
Upgrade/compatibility impact
Should not affect compatibility. Stripping of the LTO bytecode is critical to ensure there are not long term compatibility issues.
How To Test
In the short term, I'm happy to expose a repository with a gcc-10 snapshot and updated redhat-rpm-config. Developers could then use that repo to pick up gcc-10 and LTO optimizations for testing purposes. I'm already doing this internally for x86_64 and exposing it to the world would be trivial.
Given such a repository, another developer would merely use that repo when building their package. No special hardware is needed. The most useful testing is first to identify FTBFS issues and get them proactively fixed. I'm happy to own that since I'm already doing that for baseline gcc-10 issues as well as gcc-10 + LTO issues.
Doing the same testing on other architectures would definitely be useful. I'd be particularly concerned about large packages on the 32bit architectures. I wouldn't be surprised if we find some packages need to opt-out of LTO because they run out of memory at link/compile time. I'm already in contact with some Debian maintainers who want to do testing around this issue as they're investigating a similar change for Debian.
I'm already building all of Fedora with the weekly gcc-10 snapshots (including LTO builds starting the week of 12/15). This is primarily to proactive find/address issues with the gcc-10 transition, but verification of LTO state pretty much piggy backs for free on the gcc-10 work.
User Experience
In theory, the only noticeable difference to users would be smaller, faster binaries and DSOs. However, a developer that uses rpmbuild to build their own code may see their package fail to build if it's got errors or uses certain features that do not work with LTO.
Users who try to debug Fedora shipped executables could notice differences in the debugging experience.
Dependencies
None expected beyond addressing FTBFS issues and coordination between GCC and GDB teams on any debugging issues we find over the next few weeks.
Contingency Plan
- Contingency mechanism: Revert the LTO flags injection
- Contingency deadline: Beta freeze, but shooting for prior to mass rebuilds starting
- Blocks release? No
- Blocks product? No
Most critically, if we don't address the GDB testsuite issue noted above, our fallback position would be to simply disable the LTO injection globally and re-evaluate for Fedora 33, similarly if we were to find some show-stopping LTO issue.
Otherwise the plan is to analyze the remaining 100-125 package build failures. These are likely a mixture of configure issues that can't be trivially fixed via %configure, new diagnostics exposed by the deeper analysis from LTO, and other small issues.
Documentation
I would think we would want documentation on the opt-out method for RPM builds.