[LLVMbugs] [Bug 15247] New: Possible regression of vectorization of Linpack on ARM

[bugzilla-daemon at llvm.org]

http://llvm.org/bugs/show_bug.cgi?id=15247

            Bug ID: 15247
           Summary: Possible regression of vectorization of Linpack on ARM
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Loop Optimizer
          Assignee: unassignedbugs at nondot.org
          Reporter: renato.golin at linaro.org
                CC: llvmbugs at cs.uiuc.edu
    Classification: Unclassified

Created attachment 10004
  --> http://llvm.org/bugs/attachment.cgi?id=10004&action=edit
result of the runs with clang and GCC

Running Linpack from (http://www.netlib.org/benchmark/linpackc.new) with clang
with and without vectorization, I get 40% of performance boost on Intel but -3%
on ARM.

Looking at the code, the intermixing of ARM and NEON code looks particularly
bad, so it seems some shuffles are not being lowered correctly or are being
generated on a pattern that the back-end instruction combine doesn't recognize.

Nadav said he got 40% boost on ARM too, so this could be a regression, but not
necessarily on the loop vectorizer, since the BB vectorizer uses the same cost
model and is known to abuse of shuffles.

Attached is the CSV with the result of the runs with clang and GCC. The
compilation command lines are below:

GCC:
$ gcc -O3 linpack.c

Clang Vect:
$ clang -O3 linpack.c

Clang No-vect:
$ clang -O3 -fno-vectorize linpack.c

-- 
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/20130212/16441021/attachment.html>