[llvm-bugs] [Bug 26314] New: Vectorizer failure

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=26314

            Bug ID: 26314
           Summary: Vectorizer failure
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: Loop Optimizer
          Assignee: unassignedbugs at nondot.org
          Reporter: egor.kochetov at intel.com
                CC: llvm-bugs at lists.llvm.org
    Classification: Unclassified

Created attachment 15718
  --> https://llvm.org/bugs/attachment.cgi?id=15718&action=edit
The reproducer (20 SLOC, either C or C++, does not matter)

While compiling the attached 'main.cpp' I've encountered clang vectorization
problem, the performance difference is 58ms vs 20ms (almost 200%). After the
commit r257134 clang fails to vectorize 'slow_function'.

I compile with:
    clang++ -O2 -m64 -march=slm -mfpmath=sse -fPIE -flto -g -fvectorize
Results for -O2 and -Ofast, for -m32 and -m64 are all the same.

Here is what I get from an old (fast) clang:

    main.cpp:11:9: remark: the cost-model indicates that interleaving is not
beneficial [-Rpass-analysis=loop-vectorize]
            for (int k=0; k<Z; k++) {

Here is what I get from a new (slow) clang:

    main.cpp:11:9: remark: loop not vectorized: cannot check memory
dependencies at runtime [-Rpass-analysis=loop-vectorize]
            for (int k=0; k<Z; k++) {

The triggering SVN change is:

    Index: lib/Analysis/LoopAccessAnalysis.cpp
    ===================================================================
    --- lib/Analysis/LoopAccessAnalysis.cpp (revision 257133)
    +++ lib/Analysis/LoopAccessAnalysis.cpp (revision 257134)
    @@ -845,6 +845,7 @@
       if (Lp != AR->getLoop()) {
         DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop "
<<
               *Ptr << " SCEV: " << *PtrScev << "\n");
    +    return 0;
       }

       // The address calculation must not wrap. Otherwise, a dependence could
be

For comparison, here is what I get from GCC 6.0 trunk with
'-fopt-info-vec-note':

    loop at main.cpp:12: if (ivtmp_126 >= bnd.8_75)

    main.cpp:11:24: note: LOOP VECTORIZED

The binary compiled by GCC is slightly faster than that compiled by old clang
(well, if I add fwrite of s to /dev/null at the end of 'main').

-- 
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/20160126/84e17f76/attachment-0001.html>