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

[Amara Emerson via Phabricator via llvm-commits]

aemerson added a comment.

In D109858#3005799 <https://reviews.llvm.org/D109858#3005799>, @foad wrote:

> What is the compile time impact of this vs D109154 <https://reviews.llvm.org/D109154>? I'm assuming they are both equally effective at cleaning up the dead code.
>
> If we're going to go with this approach, should we clean up legalizeMachineFunction by ripping out all the bottom-up RPO order stuff, and the on-the-fly (presumably expensive) isTriviallyDead calls?

CTMark results for this patch AArch64 using 5 runs:

  O0 -g:
  
  Program             base   full-dce diff 
   pairlocalalign       1.56   1.57    1.0%
   kc                   8.70   8.70    0.0%
   bullet              14.03  14.02   -0.1%
   lencod               3.03   3.02   -0.3%
   clamscan             3.81   3.80   -0.3%
   consumer-typeset     2.79   2.78   -0.3%
   tramp3d-v4           5.71   5.69   -0.3%
   SPASS                3.37   3.36   -0.4%
   sqlite3              1.08   1.08   -0.4%
   7zip-benchmark      18.00  17.91   -0.4%
   Geomean difference                 -0.2%
  
  Os -g:
  
  Program             base   full-dce diff 
   bullet              23.96  23.99    0.1%
   lencod              10.03  10.04    0.1%
   tramp3d-v4          16.85  16.84   -0.0%
   7zip-benchmark      30.93  30.90   -0.1%
   kc                  14.95  14.94   -0.1%
   SPASS               11.03  11.01   -0.2%
   clamscan            10.71  10.68   -0.2%
   pairlocalalign       5.80   5.79   -0.3%
   sqlite3              6.75   6.72   -0.4%
   consumer-typeset     8.46   8.39   -0.8%
   Geomean difference                 -0.2%

So we actually see a minor improvement by cleaning up. Not sure why that is.

As for D109154 <https://reviews.llvm.org/D109154>:

  -O0 -g:
  Program             base   partial-dce diff 
   sqlite3              1.08   1.09       0.1%
   bullet              14.03  14.04       0.1%
   kc                   8.70   8.68      -0.2%
   pairlocalalign       1.56   1.55      -0.2%
   7zip-benchmark      18.00  17.95      -0.3%
   lencod               3.03   3.02      -0.4%
   clamscan             3.81   3.78      -0.6%
   SPASS                3.37   3.35      -0.6%
   tramp3d-v4           5.71   5.67      -0.7%
   consumer-typeset     2.79   2.76      -1.1%
   Geomean difference                    -0.4%
  
  -Os -g:
  Program             base   partial-dce diff 
   SPASS               11.03  11.03       0.0%
   bullet              23.96  23.97       0.0%
   sqlite3              6.75   6.75      -0.1%
   7zip-benchmark      30.93  30.90      -0.1%
   tramp3d-v4          16.85  16.82      -0.1%
   consumer-typeset     8.46   8.43      -0.2%
   pairlocalalign       5.80   5.79      -0.3%
   kc                  14.95  14.90      -0.3%
   lencod              10.03   9.99      -0.4%
   clamscan            10.71  10.66      -0.5%
   Geomean difference                    -0.2%

Same thing for D109154 <https://reviews.llvm.org/D109154>, but a slightly bigger improvement at -O0.


Repository:
  rG LLVM Github Monorepo

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

https://reviews.llvm.org/D109858