[LLVMbugs] [Bug 23758] New: Gigantic block frequencies produced by clang in PGO

[bugzilla-daemon at llvm.org]

https://llvm.org/bugs/show_bug.cgi?id=23758

            Bug ID: 23758
           Summary: Gigantic block frequencies produced by clang in PGO
           Product: clang
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: normal
          Priority: P
         Component: LLVM Codegen
          Assignee: unassignedclangbugs at nondot.org
          Reporter: congh at google.com
                CC: llvmbugs at cs.uiuc.edu
    Classification: Unclassified

Created attachment 14434
  --> https://llvm.org/bugs/attachment.cgi?id=14434&action=edit
CFG with edge frequencies.

When compiling spec2000 in PGO mode, I noticed many very large block
frequencies for some programs. They are too big to be true. One example is
164.gzip in spec2000. I made the profile data from running the binary with
options "input.random 60". The produced CFG is attached here, which shows the
hottest block has the frequency over 14E12. Note that the frequency of the
entry node is only 17, which means that the hottest node runs almost 1E12 times
more that the entry.

To find out the actual frequency of that hot node, I profiled the same program
using GCC, and also inserted a counter manually to the program. Both showed
that the hottest node ran only about 4E7 times more than the entry.

So I think there may be something wrong in LLVM's frequency calculation. 

(In the attached CFG, the numbers on edges are frequencies.)

-- 
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/20150604/0e7b4b38/attachment.html>