From mboxrd@z Thu Jan 1 00:00:00 1970 From: Mark Mitchell To: gcc@gcc.gnu.org Subject: Important: Development Plan for Future Releases Date: Mon, 16 Jul 2001 21:03:00 -0000 Message-id: <163530000.995341410@warlock.codesourcery.com> X-SW-Source: 2001-07/msg01178.html The Steering Committee has, after much deliberation, and after considering the various opinions presented on the public mailing lists, adopted the proposal detailed below. This policy will govern the development of GCC henceforth; it is important that all developers understand the policy. As Bernd pointed out, any policy that could be mechanically enforced is too simplistic. Therefore, there are aspects of this policy that are under-specified. We will address gray areas on a case-by-case basis. There are no absolutes; exceptions can be made by the SC to any of these provisions. And, of course, we understand that this policy is very likely to require revisions as we go forward. Although many parts of this policy follow the outline I presented some weeks back, changes have been made in many places, and, some very important goals, such as support for more targets, have been made a bigger focus. I would like to thank everyone for their contributions to the debate, and especially (in alphabetical order) Per Bothner, Joe Buck, Robert Dewar, David Edelsohn, Richard Kenner, H.J. Lu, Joseph Meyers, Alexandre Oliva, Gerald Pfeifer, Bernd Schmidt, and Joel Sherill for taking the time to make very substantial contributions. Thank you in advance for your willingness to go along with this experiment. I hope and believe that it will work out well. If it does not, we will try something different! So as to avoid surprising anyone, the first Stage 1 (see below) period will begin August 15th, 2001. Therefore, GCC 3.1 will be released on or about April 15, 2002. In the meantime, we will continue to make bug-fix releases based on GCC 3.0.x. -- I. Objectives The goals of this policy are: - Higher-quality releases - Support for more targets - Continued encouragement of major infrastructure improvements - More predictable release schedules - More frequent releases Rationale: It has been difficult for us to make consistent, high-quality releases that support a wide variety of targets. In particular, GCC 3.0 achieved a high standard of quality on many targets, but was by no means perfect, and failed to support as many targets as we would have liked. In addition, the release was late relative to original scheduling estimates. And, the time between the GCC 2.95 and GCC 3.0 releases was longer than everyone would have liked. We think that we will better serve the user community by making releases somewhat more frequently, and on a consistent schedule. In addition, a consistent schedule will make it possible for the Release Manager to better understand his or her time commitment will be when he or she agrees to take the job. II. Development Methodology Branches: The development of major changes shall be done on a branch. (Under some circumstances, development may be done in a private development tree, but the Steering Committee strongly encourages the use of a branch in the publicly accessible GCC development tree.) Examples of major changes include: - A new C++ parser - A new memory allocation scheme for internal use in the compiler - A new optimization pass There is no firm guideline for what constitutes a "major change" and what does not. If a developer is unsure, she should ask for guidance on the GCC mailing lists. In general, a change that has the potential to be extremely destabilizing should be done on a branch. Changes may be merged from a branch only after: - They meet the standards for any other check-in. For example, the code must be well-documented, and any user-visible changes, including command-line options, should be documented in the manual. - The branch has been validated on three different targets, each targeting a different microprocessor family. Validation should consist of bootstrapping the compiler (unless that is impossible for the microprocessor selected) and checking that there are no new regression test failures. It is acceptable to use a simulator for validation; the use of real hardware is not required. Rationale: Making major improvements to the compiler is good. However, such changes are inevitably somewhat destabilizing. By using a branch, it is possible for people to test and improve the code, but without impacting other developers. By testing major changes more thoroughly than ordinary changes, we will help reduce the impact on other developers. In addition, the presence of half-finished projects on the mainline presents a variety of problems when preparing for a release. If serious bugs remain, it may be impossible to do a release until the bugs are fixed, thereby impacting the release schedule. If functionality is present, but not robust (for example, command-line options that activate an optimization pass that does not work very well), then we can seriously confuse users. Schedule: Development on our main branch will proceed in three stages. Each stage will be two months in length. Stage 1 ------- During this period, changes of any nature may be made to the compiler. In particular, major changes may be merged from branches. Stage 2 ------- During this period, major changes may not be merged from branches. However, other smaller improvements may be made. For example, support for a new language construct might be added in a front-end, or support for a new variant of an existing microprocessor might be added to a back-end. Stage 3 ------- During this period, the only changes that may be made to the compiler are changes that fix bugs. New functionality may not be introduced during this period. Rationale: In order to produce releases on a regular schedule, we must ensure that the mainline is reasonably stable some time before we make the release. Therefore, more radical changes must be made earlier in the cycle, so that we have time to fix any problems that result. In order to reach higher standards of quality, we must focus on fixing bugs; by working exclusively on bug-fixing through Stage 3, we will have a higher quality source base as we prepare for a release. Although maintaining a branch, including merging new changes from the mainline, is somewhat burdensome, the absolute worst case is that such a branch will have to be maintained for four months. During two of those months, the only mainline changes will be bug-fixes, so it is unlikely that many conflicts will occur. Patch Reversion: If a patch is committed which introduces a regression [1], on any target which the Steering Committee considers to be important [2], and the problem is reported to the original poster, and 48 hours pass without either the original poster or any other party indicating that a fix will be forthcoming in the very near future, and two people with write privileges to the affected area of the compiler determine that the best course of action is to revert the patch, then they may revert the patch. After the patch has been reverted, the poster may appeal the decision to the Steering Committee. Rationale: If an important platform is broken, then it will be difficult to prepare a release. If nobody volunteers to fix the problem, then we will have an unpleasant choice: delay the release, or release a compiler that we know to be worse than the previous release. Therefore, it is important that we be able to revert patches that cause problems. In addition, regressions on the mainline can impede the development of other improvements. On the other hand, we want to encourage new development, and new development inevitable introduces new bugs. Therefore, it is important that decisions be made on a case-by-case basis, and that the proponent of a change have an opportunity to argue that the benefits of the change outweigh the costs. Therefore, the decision to revert requires two consenting parties, and such decisions may be appealed to the Steering Committee. However, during the appeal, the mainline will remain working, to avoid impeding other development. III. Release Methodology Schedule: At the conclusion of Stage 3 above, a release branch will be created. On the release branch, the focus will be fixing any regressions from the previous release, so that each release is better than the one before. The release will occur two months after the creation of the branch. (Stage 1 of the next release cycle will occur in parallel.) If, however, support for an important platform has regressed significantly from the previous release or support for a platform with active maintenance has regressed significantly relative to an earlier Stage in the current release cycle, the release will be postponed until the regressions are corrected, unless the Steering Committee releases the automatic hold on the release. Rationale: While making releases on schedule is important, quality should not be subordinate to schedule. If a major platform is not adequately supported, but was well supported previously, then we should address the problems. Presumably, this will not be unduly difficult, since we will have spent four months fixing bugs by the time the release would occur. Bug-fix Releases: Bug-fix releases will be made two and four months after the initial release, as necessary. Rationale: By waiting for two months to make a bug-fix release, we will be able to accumulate fixes for the most important problems, and avoid spending undue amounts of time on release testing and packaging. Footnotes --------- [1] Note that no distinction is made between patches which are themselves buggy and patches which expose "latent" bugs elsewhere in the compiler. [2] The Steering Committee is planning to prepare a list of such targets. The list will be revised at the beginning of each release cycle, if necessary.