Hi, After being faced with the prospect of debugging Yet Another PGO-only Bug, I propose that we stop testing Linux PGO builds. One of these bugs is causing perma-red on Aurora (bug 799295). The developer cost of figuring out how to reproduce, debug, and fix these problems is quite high, and even higher where a specific distro and compiler version might be in play. The cost is not worth the benefit of testing this configuration.
All evidence at Mozilla suggests that PGO is endlessly buggy, and I've heard this has been true historically in compilers. PGO deeply and zealously modifies calling conventions, function bodies, and large callgraphs, causing subtle problems, often topcrashes, that are difficult to reproduce and nearly impossible to debug. The result is that we litter the codebase with untestable deoptimization hints that may or may not randomly become obsolete. This is depressing on any platform, but at least on Windows, we have a single compiler version to worry about (and, notably, a compiler that has PGO bugs even with a huge amount of resources behind its development). It is possible that PGO could expose real bugs. If so, we'll see the bug one way or another, likely in a context that is more easily reproducible and easier to debug. So, let's turn off Linux PGO testing, free up releng resources, and save ourselves the inevitable pain and suffering. -David _______________________________________________ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
