[PATCH] D109858: [GlobalISel][AMDGPU] Add dead code elimination clean up after legalization.

[Amara Emerson via Phabricator via llvm-commits]

aemerson added a comment.

I measured the CT when we only delete instructions at the end, and instead of doing an RPOT walk we just process the blocks in sequential order:

  O0 -g:
  Program             base   final-dce-only diff 
   bullet              14.03  14.06          0.2%
   sqlite3              1.08   1.09          0.1%
   lencod               3.03   3.03          0.1%
   tramp3d-v4           5.71   5.71         -0.0%
   clamscan             3.81   3.80         -0.2%
   kc                   8.70   8.68         -0.3%
   7zip-benchmark      18.00  17.94         -0.3%
   SPASS                3.37   3.36         -0.4%
   pairlocalalign       1.56   1.55         -0.5%
   consumer-typeset     2.79   2.77         -0.7%
   Geomean difference                       -0.2%
  
  -Os -g:
  Program             base   final-dce-only diff 
   SPASS               11.03  11.08          0.4%
   sqlite3              6.75   6.77          0.3%
   clamscan            10.71  10.72          0.2%
   pairlocalalign       5.80   5.81          0.1%
   bullet              23.96  23.98          0.1%
   lencod              10.03  10.04          0.0%
   7zip-benchmark      30.93  30.94          0.0%
   tramp3d-v4          16.85  16.81         -0.2%
   kc                  14.95  14.89         -0.4%
   consumer-typeset     8.46   8.41         -0.5%
   Geomean difference                        0.0%

It looks like a wash, a little bit worse than before.

Overall, I think the data is showing that D109154 <https://reviews.llvm.org/D109154> is most preferable for compile time improvements, especially at -O0 where compile time is more valuable. @Petar.Avramovic let's move ahead with your patch then?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D109858/new/

https://reviews.llvm.org/D109858