[LLVMbugs] [Bug 15222] New: pcsa.f and csa.f from xplor-nih compile ~100x too slowly using -fplugin-arg-dragonegg-llvm-option=-vectorize

bugzilla-daemon at llvm.org bugzilla-daemon at llvm.org
Sat Feb 9 10:55:33 PST 2013


http://llvm.org/bugs/show_bug.cgi?id=15222

            Bug ID: 15222
           Summary: pcsa.f and csa.f from xplor-nih compile ~100x too
                    slowly using
                    -fplugin-arg-dragonegg-llvm-option=-vectorize
           Product: dragonegg
           Version: trunk
          Hardware: Macintosh
                OS: MacOS X
            Status: NEW
          Severity: normal
          Priority: P
         Component: New Bugs
          Assignee: baldrick at free.fr
          Reporter: howarth at nitro.med.uc.edu
                CC: llvmbugs at cs.uiuc.edu
    Classification: Unclassified

Created attachment 9989
  --> http://llvm.org/bugs/attachment.cgi?id=9989&action=edit
test cases exposing the vectorization bottleneck

The attached code from xplor-nih takes an abnormally long time to compile with
-fplugin-arg-dragonegg-llvm-option=-vectorize under dragonegg...

% time de-gfortran47 -c pcsa.f -Ofast
-fplugin-arg-dragonegg-llvm-option=-vectorize -fdefault-integer-8
-ffixed-line-length-120 -fno-second-underscore
80.241u 0.237s 1:20.49 99.9%    0+0k 0+22io 0pf+0w

% time de-gfortran47 -c csa.f -Ofast
-fplugin-arg-dragonegg-llvm-option=-vectorize -fdefault-integer-8
-ffixed-line-length-120 -fno-second-underscore
285.138u 0.573s 4:45.73 99.9%   0+0k 0+13io 0pf+0w

compared to the normal compilation without vectorization...

% time de-gfortran47 -c pcsa.f -Ofast -fdefault-integer-8
-ffixed-line-length-120 -fno-second-underscore
0.560u 0.043s 0:00.60 100.0%    0+0k 0+20io 0pf+0w

% time de-gfortran47 -c csa.f -Ofast -fdefault-integer-8
-ffixed-line-length-120 -fno-second-underscore
0.842u 0.059s 0:00.90 98.8%     0+0k 0+10io 0pf+0w

These may be good test cases for finding bottlenecks in the current vectorizor
code. Note that stock gfortran 4.7.2 compiles this same code with vectorization
very rapidly...

% time gfortran-fsf-4.7 -c pcsa.f -Ofast -fdefault-integer-8
-ffixed-line-length-120 -fno-second-underscore
1.187u 0.049s 0:01.29 94.5%    0+0k 1+20io 528pf+0w

% time gfortran-fsf-4.7 -c csa.f -Ofast -fdefault-integer-8
-ffixed-line-length-120 -fno-second-underscore
2.459u 0.083s 0:02.54 99.6%    0+0k 0+8io 0pf+0w

Tested on x86_64-apple-darwin12 with llvm/dragonegg svn at r174816.

-- 
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/20130209/b1fa206c/attachment.html>


More information about the llvm-bugs mailing list