[PATCH] D16404: [X86] Use hash table in LEA optimization pass

[Andrey Turetskiy via llvm-commits]

aturetsk added a comment.

As Daniel suggested in https://llvm.org/bugs/show_bug.cgi?id=25843 I added a hash table to store LEAs to search for them faster in removeRedundantLEAs and removeRedundantAddrCalc.
I used the same test from the bug as in http://reviews.llvm.org/D15692, but with -Oz option (to enable the pass fully).
Here are the results:

Without LEA pass:

  none$ time ./bin/clang++ -std=c++11 -S test-oz.ll -Oz -mllvm -disable-x86-lea-opt
  
  real    0m8.491s
  user    0m8.437s
  sys     0m0.052s

With old LEA pass:

  none$ time ./bin/clang++ -std=c++11 -S test-oz.ll -Oz -mllvm -enable-x86-lea-opt
  
  real    0m8.878s
  user    0m8.819s
  sys     0m0.051s

With new LEA pass (no need to pass -enable-lea-opt any more):

  none$ time ./bin/clang++ -std=c++11 -S test-oz.ll -Oz
  
  real    0m8.546s
  user    0m8.493s
  sys     0m0.051s

The difference for the current example is small, but visible. I think for the cases mentioned in the bug it will be more noticeable.


http://reviews.llvm.org/D16404