OptView2 MUC meetup slides
Download as PPTX, PDF0 likes314 views
The document presents 'optview2', a tool aimed at improving compiler optimizations for developers rather than compiler authors, featuring enhancements such as denoising optimization data and improved presentation. It discusses challenges with traditional optimization reporting, including high memory usage and non-actionable data. Future enhancements for optview2 include better filtering, reduced runtime, and support for code annotations.
![Some raw optimization data is exposed
• Clang/gcc: -Rpass
code.cc:4:25: remark: foo inlined into bar [-Rpass=inline]
• Intel: -qopt-report=[0..5]
• Microsoft (very partial):
/Qpar-report, /Qvec-report
3](https://image.slidesharecdn.com/optviewer2munich-220205190211/95/OptView2-MUC-meetup-slides-3-638.jpg)
OptView2 MUC meetup slides
- 1. OptView2 View and improve compiler optimizations Ofek Shilon Istra Research Jan 2022 @OfekShilon 1
- 2. Things I’d like the compiler to tell me: • I couldn’t inline that function …because I had only its declaration. • I’m re-evaluating these expressions on every loop iteration …because I couldn’t prove they stay fixed. • I’m reloading this variable from memory many times …because I don’t know whether unrelated code lines change it. 2
- 3. Some raw optimization data is exposed • Clang/gcc: -Rpass code.cc:4:25: remark: foo inlined into bar [-Rpass=inline] • Intel: -qopt-report=[0..5] • Microsoft (very partial): /Qpar-report, /Qvec-report 3
- 4. 4
- 5. Presentation Matters. • opt-viewer, 2016 work by Adam Nemet (Apple) and others • Focus on Clang, over Linux 5
- 6. Opt-Viewer Usage • Build with an extra clang switch: -fsave-optimization-record *.opt.yaml files are generated, by default in the obj folder. • Generate htmls: $ opt-viewer.py --output-dir <htmls folder> --source-dir <repo> <yamls folder> 6
- 8. Opt-Viewer Usage • Speed up yaml generation ~x50 by installing libyaml: $ sudo apt install libyaml-dev $ pip --no-cache-dir install --verbose --force-reinstall -I pyyaml • The script would detect missing libyaml and suggest installing it: “For faster parsing, you may want to install libyaml for PyYAL” 8
- 9. 9
- 10. Great work, still not much traction https://www.youtube.com/watch?v=qq0q1hfzidg 10
- 11. Great work, still not much traction 11
- 12. Why? • Heavy • High I/O • High memory • >1G htmls • Designed (and presented) for compiler authors • Mostly non actionable to developers 12
- 14. Target Developers, Not Compiler Authors • Denoise: • Collect only optimization failures • Ignore system headers • Display a single entry per type/source loc (in index) • … • ~1.5M lines ==> 22K lines • Don’t mention ‘passes’ • Make the index sortable, resizable & pageable • Abridged function names • ... 14
- 15. Example - OpenCV • file://wsl.localhost/Ubuntu-20.04/home/ofek/html/opencv- clean/modules/photo/index.html 15
- 17. ! 17
- 18. “Clobbered by store” • https://godbolt.org/z/K9a79MhcE • Aliasing • Forces reloading redistBatch on every loop iteration 18
- 19. Aliasing – the silent killer • Distribution of 22K remarks in a C++ project: 19
- 20. 20
- 21. “Clobbered by call” • https://godbolt.org/z/bb9fTKj9d • https://github.com/llvm/llvm-project/issues/53102 21
- 22. 22
- 23. Aliasing: Counter-measures • LTO ? • __restrict__ • Non standard, tells the compiler an argument doesn’t alias with anything else • __attribute__((pure)), __attribute__((const)) • Non standard, tell the compiler a function doesn’t write/read global state. • Potentially resolve ‘clobbered by call’ opt failures, • Strict-Aliasing • The only C++ standard compliant way. 23
- 24. Strict Aliasing • “Strict aliasing is an assumption made by the compiler, that objects of different types will never refer to the same memory location (i.e. alias each other.)” Mike Acton https://cellperformance.beyond3d.com/articles/2006/06/understanding-strict-aliasing.html • Allowed optimization. on by default for -O2 + • char* is an exception – always allowed to alias with any type. • Can be disabled with –fno-strict-aliasing (pessimization!) 24
- 25. Strict Aliasing: Forcing type difference • typedef / using ? • Inherit int? • “I considered …to allow derivation from built-in classes … However, I restrained myself. … the C conversion rules are so chaotic that pretending that int, short, etc., are well-behaved ordinary classes is not going to work. They are either C compatible, or they obey the relatively well-behaved C++ rules for classes, but not both.” Bjarne Stroustrup, The Design and Evolution of C++, §15.11.3. • Strong-Typedefs • Wrappers 25
- 26. Strict Aliasing demos • https://godbolt.org/z/qGdj4qoG1 • https://godbolt.org/z/zWfYdsorE • https://godbolt.org/z/oxq1K3n8W 26
- 27. Demo 3: loop hoisting 27
- 28. • Check foreach!! • Try to form a toy example 28
- 29. Demo 3: loop hoisting 29
- 30. 30
- 31. GCC work • https://gcc.gnu.org/legacy-ml/gcc-patches/2018- 05/msg01675.html • https://github.com/davidmalcolm/gcc-opt-viewer • YAML -> JSON.gz • Actually a better choice https://stackoverflow.com/questions/27743711/can-i-speedup-yaml • Active only during 2018 • Currently seems unusable https://github.com/davidmalcolm/gcc-opt-viewer/issues/3 31
- 32. Conclusions applicable to other compilers? • In many cases – yes. • Shows where to look 32
- 33. OptView2 Future work • Improve filtering, • Reduce run time & memory, • Run on windows, • Consume binary optimization remarks, • Revive (possibly integrate) gcc-opt-viewer, • Support code annotations, • Report LLVM bugs. 33
- 34. Lots of help needed • https://github.com/Of ekShilon/optview2 • ofekshilon@gmail.com 34
- 35. Bkp 35
- 36. llvm-opt-report $ clang -O3 -o /tmp/v.o -c /tmp/v.c -fsave-optimization-record $ llvm-opt-report /tmp/v.yaml > /tmp/v.lst $ cat /tmp/v.lst • https://reviews.llvm.org/D25262 • https://github.com/llvm/llvm- project/tree/main/llvm/tools/llvm-opt- report 36
- 37. Impact https://github.com/opencv/opencv/wiki/HowToUsePerfTests $ cd ~/src/opencv/modules/ts/misc/ $ python3 ./run.py ~/src/opencv-clean/build/ -w ~/logs-opencv-clean $ python3 ./run.py ~/src/opencv-opt/build/ -w ~/logs-opencv-opt $ python3 ./summary.py ~/logs-opencv-clean/core* ~/logs-opencv-opt/core* -u mks -m median • Seems substantial. • Will publish more detailed steps in the optview2 github page (and possibly fork opencv) Detour 37
- 38. 38
Editor's Notes
- #2: Very little known LLVM feature, I really think can add value to your day to day work. Certainly if you’re working in clang I’ll present first the LLVM existing opt-viewer tool, then optview2 which is derived work by me, And about half the time would be dedicated to demonstrations of this tool in use. With specific focus on aliasing issues.
- #3: I want to gain visibility into compiler optimizations. More specifically, I want to see where the compiler tried optimizations and failed – there might be something I could do about it.
- #7: LLVM was already instrumented to show optimization choices (for –Rpass), Hal Finkel wrapped the data in yaml format for machine consumption. Adam Nemet wrote the opt-viewer python scripts
- #8: A few small python scripts that process the compiler-emitted remarks into annotated source htmls, and an index page – which is an optimization dashboard of sorts.
- #13: Don’t know, guessing I stopped it at 60G mem Ran it on separate subsets of a real project An extra step is needed to make this usable to us Let’s take that extra step.
- #16: The demos are of results of the optview2 version, but let it be said that the entire heavy lifting is done by the original code Credit to Ilan Ben Hagai for javascript wizardry
- #18: Full opt-viewer, not just failures
- #19: file:///C:/optview2/html/opencv4/modules_imgproc_src_clahe.cpp.html#L201 One of the darker corners corners of C++, already full of dark corners Some guesswork in interpreting the optview2 reports
- #24: Restrict –compiles on non arguments, didn’t see any impact
- #25: Turning on strict-aliasing means allowing the compiler to assume exactly that. Turning it off is a pessimization What can be done? Force type difference
- #28: file:///C:/optview2/html/opencv2/_home_ofek_src_opencv_modules_imgproc_src_canny.cpp.html#L419
- #30: This is not good C++ code Don’t do this for fun. Do this in code where you *really* care about cycle-level performance. Hopefully compilers would get good enough to not have to do this.
- #33: Probably If your project *can* be built by clang – I suggest you give it a go.
- #34: Report LLVM bugs: alias analysis bugs had no observable symptoms until now. There are no diagnostics emitted on them and they don’t result in bad codegen So, I suspect there are plenty of them.
- #38: Tests done on WSL2, sys calls have overhead – but in both repos