[llvm-bugs] [Bug 45679] New: LoopVectorizer produces wrong code with UserVF not dividing small TripCount

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=45679

            Bug ID: 45679
           Summary: LoopVectorizer produces wrong code with UserVF not
                    dividing small TripCount
           Product: libraries
           Version: trunk
          Hardware: All
                OS: All
            Status: NEW
          Severity: normal
          Priority: P
         Component: Loop Optimizer
          Assignee: unassignedbugs at nondot.org
          Reporter: ayal.zaks at intel.com
                CC: dorit.nuzman at intel.com, florian_hahn at apple.com,
                    gil.rapaport at intel.com, llvm-bugs at lists.llvm.org

Created attachment 23402
  --> https://bugs.llvm.org/attachment.cgi?id=23402&action=edit
Reproducer

Compiling the attached with

  opt -loop-vectorize -force-vector-width=4 -S

vectorizes a loop with trip-count=14 using VF=4 w/o a tail, explaining that:

LV: Found trip count: 14
LV: Not allowing scalar epilogue due to low trip count.
LV: Performing code size checks.
LV: The Smallest and Widest types: 32 / 32 bits.
LV: The Widest register safe to use is: 32 bits.
LV: No tail will remain for any chosen VF.  <== but LV doesn't get to chose VF
LV: Using user VF 4.


Culprit is computeMaxVF() who checks if TC % MaxVF == 0 where MaxVF =
computeFeasibleMaxVF(TC) (=1 in this case), even if a UserVF is given and
ultimately used.

-- 
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/20200426/30773211/attachment.html>