[llvm-bugs] [Bug 39852] New: [loop vectorize] failure to vectorize with llvm.invariant.start after instcombine

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=39852

            Bug ID: 39852
           Summary: [loop vectorize] failure to vectorize with
                    llvm.invariant.start after instcombine
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Loop Optimizer
          Assignee: unassignedbugs at nondot.org
          Reporter: mpfergu at gmail.com
                CC: llvm-bugs at lists.llvm.org

Created attachment 21183
  --> https://bugs.llvm.org/attachment.cgi?id=21183&action=edit
invariant-start-vector-issue.ll

I'm seeing an issue where adding !llvm.invariant.start metadata seems to
prevents loop vectorization from occurring.

I'd like the attached example to vectorize, but it doesn't. To reproduce:

  opt -S -O3 invariant-start-vector-issue.ll -force-vector-width=4
-force-vector-interleave=1

and notice that the fmul is not vectorized. (invariant-start-vector-issue.ll is
attached). If the llvm.invariant.start metadata is removed (or they are updated
to all make %14 invariant, for example), then the loop is vectorized.

This is very surprising. I think most of us expect that adding additional
information to the LLVM IR, like llvm.invariant.start, will help and not hinder
optimization.

I know a little bit more about what is happening here. instcombine is creating
a bitcast that seems to trip up the loop vectorizer.

 opt -instcombine -S invariant-start-vector-issue.ll

generates this snippet:

  %44 = bitcast double* %4 to i64*
  %45 = load i64, i64* %44, align 8, !tbaa !2
  %46 = bitcast double* %31 to i64*
  store i64 %45, i64* %46, align 8, !tbaa !2

I believe this snippet is tripping up the LoopVectorizer.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20181130/7a2d3f1a/attachment.html>