[LLVMdev] Analysis of polly-detect overhead in oggenc

[Star Tan]

At 2013-07-14 23:30:31,"Sebastian Pop" <sebpop at gmail.com> wrote:
>On Sun, Jul 14, 2013 at 10:17 AM, Star Tan <tanmx_star at yeah.net> wrote:
>> Hi Sebastian,
>>
>> Yes, you have pointed an important reason. If we comment this source code
>> you have listed, then the compile-time overhead for oggenc*8.ll can be
>> reduced from 40.5261 ( 51.2%) to 20.3100 ( 35.7%).
>>
>> I just sent another mail to explain why polly-detect pass leads to
>> significant compile-time overhead. Besides the reason you have pointed,
>> another reason is resulted from those string buffer operations in "INVALID"
>> MACRO. If we comment both the string buffer operations in "INVALID" MACRO
>> and in the "isValidMemoryAccess" function, the compile-time overhead for
>> oggenc*8.ll would be reduced from 40.5261 ( 51.2%) to 5.8813s (15.9%).
>
>Awesome, thanks for the analysis.
>
>Can you run again perf on the resulting program: I would still like to
>understand
>where we spend the 5.88s in the rest of scop detection.
>
The top ten functions reported by perf are:
+  12.68%  opt  [kernel.kallsyms]      [k] 0xc111472c 
+   9.40%  opt  libc-2.17.so           [.] 0x0007875f
+   2.98%  opt  opt                    [.] __x86.get_pc_thunk.bx 
+   2.42%  opt  [vdso]                 [.] 0x00000425 
+   1.46%  opt  libgmp.so.10.0.5       [.] __gmpn_copyi_x86 
+   1.11%  opt  libc-2.17.so           [.] free 
+   1.07%  opt  opt                    [.] bool llvm::DenseMapBase<llvm::DenseMap<void const*, llvm::Pass*, llvm::DenseMapInfo<void const*>
+   1.02%  opt  opt                    [.] llvm::ComputeMaskedBits(llvm::Value*, llvm::APInt&, llvm::APInt&, llvm::DataLayout const*, unsigna
+   1.00%  opt  opt                    [.] llvm::Use::getImpliedUser() const 
+   0.90%  opt  libgmp.so.10.0.5       [.] __gmpz_set 
+   0.76%  opt  opt                    [.] llvm::SmallPtrSetImpl::insert_imp(void const*) 
+   0.74%  opt  opt                    [.] llvm::InstCombiner::DoOneIteration(llvm::Function&, unsigned int)
+   0.73%  opt  opt                    [.] llvm::PassRegistry::getPassInfo(void const*) const
+   0.72%  opt  libc-2.17.so           [.] malloc   
+   0.72%  opt  opt                    [.] llvm::TimeRecord::getCurrentTime(bool) 
+   0.71%  opt  opt                    [.] llvm::ValueHandleBase::AddToUseList() 
+   0.57%  opt  opt                    [.] llvm::SlotTracker::processModule() 
+   0.52%  opt  opt                    [.] llvm::PMTopLevelManager::findAnalysisPass(void const*)  
+   0.51%  opt  opt                    [.] llvm::APInt::~APInt()    
+   0.51%  opt  libgmp.so.10.0.5       [.] __gmpz_mul 

Unfortunately, I cannot set breakpoints for the top 2 functions. 
Even with "perf -g", I still cannot track where time is spent on in Polly-detect pass. The "perf -g" results are like this:
-  12.68%  opt  [kernel.kallsyms]      [k] 0xc111472c                                                                                       `
   - 0xc161984a                                                                                                                             a
   - 0xb7783424                                                                                                                             a
      - 99.93% 0x9ab2000                                                                                                                    a
         + 85.99% 0                                                                                                                         a
         + 3.14% 0xb46d018                                                                                                                  a
         + 1.42% 0xb7745000                                                                                                                 a
   - 0xc1612415                                                                                                                             a
      + 97.76% 0xc1062977                                                                                                                   a
      + 1.75% 0xc1495888 

Do you have some suggestions?

Best wishes,
Star Tan
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20130714/d697c503/attachment.html>