[PATCH] D26527: Use profile info to adjust loop unroll threshold.

[Dehao Chen via llvm-commits]

danielcdh added a comment.

The perf/size impact of this patch is small on speccpu as flat loop is rare in most of the benchmarks.

spec/2006/fp/C++/444.namd                 25.47  +0.30%
spec/2006/fp/C++/447.dealII               45.46  +0.23%
spec/2006/fp/C++/450.soplex               43.38  +0.58%
spec/2006/fp/C++/453.povray               37.88  -0.78%
spec/2006/fp/C/433.milc                   23.75  -0.13%
spec/2006/fp/C/470.lbm                    41.53  -0.09%
spec/2006/fp/C/482.sphinx3                48.97  -0.11%
spec/2006/int/C++/471.omnetpp             22.79  -0.22%
spec/2006/int/C++/473.astar               22.99  +0.17%
spec/2006/int/C++/483.xalancbmk           38.55  -0.42%
spec/2006/int/C/400.perlbench             37.06  +1.14%
spec/2006/int/C/401.bzip2                 23.38  +0.98%
spec/2006/int/C/403.gcc                   34.52  -0.25%
spec/2006/int/C/429.mcf                   42.28  -0.05%
spec/2006/int/C/445.gobmk                 27.98  +0.60%
spec/2006/int/C/456.hmmer                 26.01  -0.06%
spec/2006/int/C/458.sjeng                 30.42  +0.50%
spec/2006/int/C/462.libquantum            57.48  +0.37%
spec/2006/int/C/464.h264ref                47.6  -0.70%

geometric mean                                   +0.11%


https://reviews.llvm.org/D26527