[llvm] [DeveloperPolicy] Add guidelines for adding/enabling passes (PR #158591)

[Alexis Engelke via llvm-commits]

================
@@ -1185,6 +1185,50 @@ Suggested disclaimer for the project README and the main project web page:
    necessarily a reflection of the completeness or stability of the code, it
    does indicate that the project is not yet endorsed as a component of LLVM.
 
+Adding or enabling a new LLVM pass
+----------------------------------
+
+The guidelines here are primarily targeted at the enablement of new major
+passes in the target-independent optimization pipeline. Small additions, or
+backend-specific passes, require a lesser degree of care.
+
+When adding a new pass, the goal should be to enable it as part of the default
+optimization pipeline as early as possible and then continue development
+incrementally. The recommended workflow is:
+
+1. Implement a basic version of the pass and add it to the pass pipeline behind
+   a flag that is disabled by default.
+2. Enable the pass by default. Separating this step allows easily disabling the
+   pass if issues are encountered, without having to revert the entire
+   implementation.
+3. Incrementally extend the pass with new functionality. As the pass is already
+   enabled, it becomes easier to identify the specific change that has caused a
+   regression in correctness, optimization quality or compile-time.
+
+When enabling a pass, regardless of whether it is old or new, certain
+requirements must be met (in no particular order):
+
+ * **Maintenance:** The pass (and any analyses it depends on) must have at
+   least one maintainer.
+ * **Usefulness:** There should be evidence that the pass improves performance
+   (or whatever metric it optimizes for) on real-world workloads. Improvements
+   seen only on synthetic benchmarks may be insufficient.
----------------
aengelke wrote:

Maybe we should consider distinguishing -O2 and -O3? We already run a lot of passes that do nothing for most functions. For -O2 (the average RelWithDebInfo build) we might want to focus on transforms that have an impact for a broader range of programs and not just a few special workloads.

https://github.com/llvm/llvm-project/pull/158591