[llvm-bugs] [Bug 26055] New: LiveDebugValues is very slow

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=26055

            Bug ID: 26055
           Summary: LiveDebugValues is very slow
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: Common Code Generator Code
          Assignee: vikram.tarikere at gmail.com
          Reporter: vonosmas at gmail.com
                CC: aprantl at apple.com, dberlin at dberlin.org,
                    dblaikie at gmail.com, llvm-bugs at lists.llvm.org
    Classification: Unclassified

LiveDebugValues pass re-committed in r255759 is slow on large functions with
lots of local variables and basic blocks. This often happens if one is using
"-O2 -fsanitize=address -g" (first flag causes a lot of inlining, while ASan
adds a number of non-trivial basic blocks).

Attached is a reproducer created from OpenCV 2.2.0:

$ ./bin/llc -filetype=obj a.ll

takes forever, and the time is spent in this pass:

 46.42%  llc  llc                  [.] (anonymous
namespace)::LiveDebugValues::VarLoc::operator==((anonymous
namespace)::LiveDebugValues::VarLoc const&) const
 20.04%  llc  llc                  [.] (anonymous
namespace)::LiveDebugValues::transfer(llvm::MachineInstr&, std::list<(anonymous
namespace)::LiveDebugValues::VarLoc, std::allocator<(anonymous
namespace)::LiveDeb
 11.14%  llc  llc                  [.] bool
__gnu_cxx::__ops::_Iter_pred<(anonymous
namespace)::LiveDebugValues::transferRegisterDef(llvm::MachineInstr&,
std::list<(anonymous namespace)::LiveDebugValues::VarLoc,
  9.74%  llc  llc                  [.] (anonymous
namespace)::LiveDebugValues::runOnMachineFunction(llvm::MachineFunction&)

Note that some functions are wildly inefficient: e.g.
transferTerminatorInst appends one list to another in O(n^2) time. It looks
like at least in some cases it would be beneficial to replace lists with hash
maps. I've created a patch to do so, but the compile time was huge anyway.

Also, do you have an estimate on the total number of times you will consider a
single machine instruction? (as you can scan the same MBB several times).

Currently there's no way to disable this pass: could you consider adding an
option to switch it on/off, or disabling it while working on speeding it up?

-- 
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/20160107/07f26bf1/attachment.html>