[PATCH] D102002: [PassManager] unify vector passes between regular and LTO pipelines

[Dave Green via Phabricator via llvm-commits]

dmgreen added a comment.

> Right now, we only have tests that show less unrolling for x86, but that doesn't actually seem like a bad thing to me.

The phase ordering test in LLVM have never been very extensive. LLVM doesn't have a lot of end-to-end tests in-tree, always leaving that to C tests from the llvm-test-suite or other benchmark suites.

Without a lot of evidence that this is better across a range of architectures, I would recommend trying to be conservative, taking things a step at a time.
We have one test where this patch would make it 88% worse. That might be a pathological case but no one is going to put up with on tenth of the speed :-)



================
Comment at: llvm/lib/Passes/PassBuilder.cpp:1199
+
+  // The vectorizer may have significantly shortened a loop body; unroll again.
+  // Unroll small loops to hide loop backedge latency and saturate any parallel
----------------
I'm not sure what this comment is trying to say exactly. I think it's coming from somewhere that is very old and out of date now.

My very high level understand of the pass pipeline, at least for non-LTO in terms of loops is that we do:
 - Cleanup clang code, code simplification, inlining, DSE, GVN, all that goodness.
 - Run loop optimizations like licm, idiom recognition, loop deletion. And including _full_ unrolling.
 - Other optimizations.
 - Run Vectorizer
 - Run SLP
 - _Runtime_ Unroll loops.

There are some extra simplifications that need to happen in between too. The last unrolling, especially on smaller inorder cores has nothing really to do with vectorization. It's done near the end of the pipeline because the runtime unrolling isn't expected to be helpful to anything else, but up to that point we have not done runtime unrolling.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D102002/new/

https://reviews.llvm.org/D102002