[PATCH] D15692: [X86] Reduce complexity of the LEA optimization pass

[Andrey Turetskiy via llvm-commits]

aturetsk added a comment.

Here are the results.

Just -Os (LEA pass is disabled):

  $ time ./bin/clang++ -std=c++11 -S a.ll -Os
  
  real    0m8.328s
  user    0m8.282s
  sys     0m0.041s

-Os with the old LEA pass:

  $ time ./bin/clang++ -std=c++11 -S a.ll -Os -mllvm -enable-x86-lea-opt
  
  real    0m11.653s
  user    0m11.591s
  sys     0m0.059s

-Os with the new LEA pass:

  $ time ./bin/clang++ -std=c++11 -S a.ll -Os -mllvm -enable-x86-lea-opt
  
  real    0m8.446s
  user    0m8.380s
  sys     0m0.064s

a.ll is taken from the example from https://llvm.org/bugs/show_bug.cgi?id=25843 and was generated this way:

  $ python a.py 5000 > a.cc
  $ ./bin/clang++ -std=c++11 -S -emit-llvm -Os a.cc
  (this took about 1.5 hours)


http://reviews.llvm.org/D15692