[PATCH] D109958: [LoopFlatten] Enable it by default

[Sjoerd Meijer via Phabricator via llvm-commits]

SjoerdMeijer created this revision.
SjoerdMeijer added reviewers: dmgreen, nikic, fhahn, lebedev.ri, alanphipps, alexey.zhikhar, efriedma.
Herald added subscribers: ormris, wenlei, steven_wu, hiraditya.
SjoerdMeijer requested review of this revision.
Herald added a project: LLVM.

LoopFlatten improves a well known embedded benchmark with highly-popular industry applications with a few percentage points. But it is not restricted to just optimise a single benchmark case. Find below results for the llvm test suite and the number of loops it flattened:

  Test                                                                       # Loops flattened
  --------------------------------------------------------------------------------------------
  MultiSource/Applications/JM/lencod/lencod                                  3
  MultiSource/Benchmarks/mediabench/jpeg/jpeg-6a/cjpeg                       1
  MultiSource/Benchmarks/MiBench/consumer-jpeg/consumer-jpeg                 3
  MultiSource/Applications/JM/ldecod/ldecod                                  1       
  MultiSource/Benchmarks/ASCI_Purple/SMG2000/smg2000                         3       
  MultiSource/Benchmarks/tramp3d-v4/tramp3d-v4                              17      
  SingleSource/Benchmarks/Misc/himenobmtxpa                                  2       
  MicroBenchmarks/ImageProcessing/AnisotropicDiffusion/AnisotropicDiffusion  2
  MicroBenchmarks/ImageProcessing/BilateralFiltering/BilateralFilter         2
  MultiSource/Benchmarks/DOE-ProxyApps-C/miniGMG/miniGMG                    20
  MultiSource/Benchmarks/Rodinia/pathfinder/pathfinder                       1
  MicroBenchmarks/ImageProcessing/Blur/blur                                  2
  MicroBenchmarks/ImageProcessing/Dither/Dither                              2
  MicroBenchmarks/ImageProcessing/Dilate/Dilate                              2
  MultiSource/Benchmarks/DOE-ProxyApps-C++/HPCCG/HPCCG                       1
  MultiSource/Benchmarks/DOE-ProxyApps-C/SimpleMOC/SimpleMOC                 1
  MicroBenchmarks/ImageProcessing/Interpolation/Interpolation                2
  MultiSource/Benchmarks/ASC_Sequoia/AMGmk/AMGmk                             2
  MultiSource/Benchmarks/Rodinia/backprop/backprop                           1
  ----------------------------------------------------------------------------------------------
  Total                                                                     68

While the implementation of LoopFlatten recognises a few patterns and could be  made more generic, I believe these numbers show that it's generic enough to trigger on a wide variety of code bases, making it worthwile to enable it by default.

LoopFlatten is a relatively simple pass, it e.g. doesn't implement a computationally expensive algorithm, and doesn't require more analysis than a
typical loop pass. Compile-times for the llvm test suite (ClamAV, 7zip, tramp3d-v4, kimwitu++, sqlite3, mafft, SPASS, lencod, Bullet) show a very minor increase of ~0.04% to 0.28%. There are cases that improve compile times, but I haven't analysed that and don't want to claim of course that in general it will improve compile-times.

We have LoopFlatten enable by default downstream for many years now, thus it should have had a lot of exposure and usage and we are not aware of any problems.


https://reviews.llvm.org/D109958

Files:
  llvm/lib/Transforms/IPO/PassManagerBuilder.cpp
  llvm/test/Other/new-pm-defaults.ll
  llvm/test/Other/new-pm-lto-defaults.ll
  llvm/test/Other/new-pm-thinlto-defaults.ll
  llvm/test/Other/new-pm-thinlto-postlink-pgo-defaults.ll
  llvm/test/Other/new-pm-thinlto-postlink-samplepgo-defaults.ll
  llvm/test/Other/new-pm-thinlto-prelink-pgo-defaults.ll
  llvm/test/Other/new-pm-thinlto-prelink-samplepgo-defaults.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D109958.373180.patch
Type: text/x-patch
Size: 5440 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210917/e18a1286/attachment.bin>