[llvm-bugs] [Bug 26214] New: IR verifier slows down linking Chrome with LTO by 2.5x

via llvm-bugs llvm-bugs at lists.llvm.org
Tue Jan 19 15:51:06 PST 2016


            Bug ID: 26214
           Summary: IR verifier slows down linking Chrome with LTO by 2.5x
           Product: new-bugs
           Version: unspecified
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: new bugs
          Assignee: unassignedbugs at nondot.org
          Reporter: krasin at google.com
                CC: llvm-bugs at lists.llvm.org
    Classification: Unclassified

IR verifier added in http://reviews.llvm.org/rL257825 and
http://reviews.llvm.org/rL257823 significantly slows down LTO build. The larger
the target, the higher ratio of the slowdown. Linking Chrome with LTO is 27
mins before the changes, and 88 minutes after them.

Here is the profile:

+  40.50%        ld.gold  LLVMgold.so          [.] bool llvm::function_ref<bool
(llvm::Value const*)>::callback_fn<(anonymous
+  19.61%        ld.gold  LLVMgold.so          [.] llvm::Use::getImpliedUser()
+   7.20%        ld.gold  LLVMgold.so          [.]
llvm::IRMover::move(llvm::Module&, llvm::ArrayRef<llvm::GlobalValue*>,
std::function<void (llvm::GlobalValue&, std::function<void
(llvm::GlobalValue&)>)>, llvm
+   2.63%        ld.gold  LLVMgold.so          [.]
llvm::SmallPtrSetImplBase::FindBucketFor(void const*) const
+   1.92%        ld.gold  [kernel.kallsyms]    [k] 0xffffffff81180fa0
+   1.60%        ld.gold  LLVMgold.so          [.] forEachUser(llvm::Value
const*, llvm::SmallPtrSet<llvm::Value const*, 32u>&, llvm::function_ref<bool
(llvm::Value const*)>) [clone .constprop.1086]
+   1.58%        ld.gold  libc-2.19.so         [.] _int_malloc
+   1.18%        ld.gold  LLVMgold.so          [.] llvm::Use::getUser() const
+   0.82%        ld.gold  LLVMgold.so          [.]
llvm::SmallPtrSetImplBase::insert_imp(void const*)                             
+   0.76%        ld.gold  LLVMgold.so          [.]
llvm::NamedMDNode::getOperand(unsigned int) const

This regression is severe for our use case (deploying Control Flow Integrity in
Chrome), but I don't think it's limited just to us. Anyone who uses LTO is

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/20160119/50cc31eb/attachment.html>

More information about the llvm-bugs mailing list