[PATCH] D23388: [LoopUnroll] By default disable unrolling when optimizing for size.
Michael Zolotukhin via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Tue May 23 15:59:03 PDT 2017
mzolotukhin added a comment.
Hi Davide,
I haven't reevaluated this. At that time, we spotted several compile-time regressions (see the table below), and my goal was simply to change the unrolling behavior back to the original. I didn't try to tune the thresholds in any way, so if anyone has a compelling data, we could definitely revisit the values.
| SingleSource/Benchmarks/Stanford/Perm | 37.37% |
| MultiSource/Benchmarks/tramp3d-v4/tramp3d-v4 | 33.95% |
| SingleSource/Regression/C/bigstack | 22.07% |
| MultiSource/Benchmarks/Olden/power/power | 13.49% |
| MultiSource/Benchmarks/FreeBench/fourinarow/fourinarow | 12.68% |
| MultiSource/Benchmarks/Olden/health/health | 8.46% |
| MultiSource/Benchmarks/Olden/bh/bh | 7.00% |
| MultiSource/Benchmarks/mediabench/g721/g721encode/encode | 5.63% |
| MultiSource/Benchmarks/nbench/nbench | 5.29% |
| External/SPEC/CINT2006/462_libquantum/462_libquantum | 5.17% |
| External/SPEC/CFP2006/447_dealII/447_dealII | 4.74% |
| MultiSource/Benchmarks/MiBench/telecomm-gsm/telecomm-gsm | 2.38% |
| MultiSource/Benchmarks/mediabench/gsm/toast/toast | 2.28% |
| External/SPEC/CFP2006/433_milc/433_milc | 2.08% |
| MultiSource/Benchmarks/MiBench/consumer-lame/consumer-lame | 1.62% |
| MultiSource/Applications/JM/lencod/lencod | 1.61% |
| MultiSource/Benchmarks/MiBench/automotive-susan/automotive-susan | 1.48% |
| MultiSource/Benchmarks/Olden/bh/bh | -1.67% |
Repository:
rL LLVM
https://reviews.llvm.org/D23388
More information about the llvm-commits
mailing list