[LLVMbugs] [Bug 13821] New: Slow compilation of ASan unittests in CMake debug build

[bugzilla-daemon at llvm.org]

http://llvm.org/bugs/show_bug.cgi?id=13821

             Bug #: 13821
           Summary: Slow compilation of ASan unittests in CMake debug
                    build
           Product: new-bugs
           Version: unspecified
          Platform: All
        OS/Version: All
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: new bugs
        AssignedTo: unassignedbugs at nondot.org
        ReportedBy: samsonov at google.com
                CC: chandlerc at gmail.com, kcc at google.com,
                    llvmbugs at cs.uiuc.edu
    Classification: Unclassified


To reproduce:
1. Create cmake build directory:
$ mkdir llvm_debug_build
$ cd llvm_debug_build
$ cmake /path/to/llvm/checkout

2. Build ASan unittests
$ make AsanUnitTests -j8
OR download asan_test_preproc.cc from attachment
$ ./llvm_debug_build/bin/clang++ -O2 -faddress-sanitizer asan_test_preproc.cc
-c -o a.o

compilation time with -O2: 1m59.702s
compilation time w/o  -O2: 0m31.462s

"perf record -g" for -O2 doesn't really look helpful:

+      1.60%    [.] llvm::PointerIntPair<llvm::IndexListEntry*, 2u, unsigned
int, llvm
+      1.23%    [.] llvm::Value::getValueID() const                             
+      0.82%    [.] llvm::IndexListEntry::getIndex() const                      
+      0.78%    [.]
llvm::PointerLikeTypeTraits<llvm::IndexListEntry*>::getFromVoidPoi
+      0.73%    [.] llvm::SlotIndex::getIndex() const                           
+      0.73%    [.] llvm::PointerIntPair<llvm::IndexListEntry*, 2u, unsigned
int, llvm
+      0.71%    [.] llvm::SlotIndex::isValid() const                            
+      0.61%    [.] _int_malloc                                                 
+      0.60%    [.] llvm::SlotIndex::listEntry() const                          
+      0.51%    [.] llvm::PointerIntPair<llvm::Value*, 2u, unsigned int,
llvm::Pointer
+      0.49%    [.] llvm::SmallPtrSetImpl::insert_imp(void const*)              
+      0.46%    [.] llvm::SlotIndex::getSlot() const                            
+      0.46%    [.] llvm::APInt::isSingleWord() const                           
+      0.45%    [.] llvm::SDNode::getOpcode() const                             
+      0.42%    [.] llvm::MachineOperand::isReg() const                         
+      0.39%    [.] llvm::SmallVectorBase::empty() const                        
+      0.37%    [.] memcpy                                                      
+      0.37%    [.] llvm::SmallPtrSetImpl::FindBucketFor(void const*) const     
+      0.34%    [.] llvm::Instruction::getOpcode() const                        
+      0.33%    [.]
llvm::SpillPlacement::Node::update(llvm::SpillPlacement::Node cons
+      0.33%    [.] llvm::Type::getTypeID() const    

It would be cool if llvm could print the time spent by each pass. Does anyone
knows if it can? :) For now hypothesis is that we can blame RegAlloc, but more
investigation needed.

The test is 2000+ lines of heavy googletest code, so it's really big, but the
long compile time is somewhat distracting anyway. Possible solution is to split
the file into several smaller ones: this can save compile time, but this
doesn't fix the source of the problem.

-- 
Configure bugmail: http://llvm.org/bugs/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.