SLPVectorizer blows up compile times on certain input

[aeubanks]

Running opt -p slp-vectorizer /tmp/h.ll -disable-output takes >1 min on my machine, and much longer on weaker build farm machines. -ftime-trace shows that this took more than half the total compile time.

h.ll.txt

[alexey-bataev]

Are you using debug build of the compiler? For me it takes 7 seconds with the release version of the compiler, even though the input file is 17MB

[aeubanks]

Double checking today, it's an optimized build with asserts at 6e73750, which is ~2 weeks ago, and it took ~100s. However, moving to ToT from today (c6cce26), it has gone down to ~9s for me, so something was improved in the last couple of weeks. However, running opt -O3 --time-passes over the provided IR, it still takes a lot more time than all the other passes, which sorta points to it potentially blowing up compile times on even larger input. It'd be nice to understand if there's a specific part of SLPVectorizer that's eating up compile times on this IR and if we can bound that part? Otherwise perhaps there's nothing left to do.

===-------------------------------------------------------------------------===
                          Pass execution timing report
===-------------------------------------------------------------------------===
  Total Execution Time: 16.2712 seconds (16.2773 wall clock)

   ---User Time---   --System Time--   --User+System--   ---Wall Time---  --- Name ---
   7.7776 ( 48.1%)   0.0159 ( 16.8%)   7.7936 ( 47.9%)   7.7962 ( 47.9%)  SLPVectorizerPass
   3.2544 ( 20.1%)   0.0039 (  4.1%)   3.2582 ( 20.0%)   3.2595 ( 20.0%)  InstCombinePass
   1.4416 (  8.9%)   0.0239 ( 25.1%)   1.4655 (  9.0%)   1.4665 (  9.0%)  GVNPass
...

[alexey-bataev]

Double checking today, it's an optimized build with asserts at 6e73750, which is ~2 weeks ago, and it took ~100s. However, moving to ToT from today (c6cce26), it has gone down to ~9s for me, so something was improved in the last couple of weeks. However, running opt -O3 --time-passes over the provided IR, it still takes a lot more time than all the other passes, which sorta points to it potentially blowing up compile times on even larger input. It'd be nice to understand if there's a specific part of SLPVectorizer that's eating up compile times on this IR and if we can bound that part? Otherwise perhaps there's nothing left to do.

===-------------------------------------------------------------------------===
                          Pass execution timing report
===-------------------------------------------------------------------------===
  Total Execution Time: 16.2712 seconds (16.2773 wall clock)

   ---User Time---   --System Time--   --User+System--   ---Wall Time---  --- Name ---
   7.7776 ( 48.1%)   0.0159 ( 16.8%)   7.7936 ( 47.9%)   7.7962 ( 47.9%)  SLPVectorizerPass
   3.2544 ( 20.1%)   0.0039 (  4.1%)   3.2582 ( 20.0%)   3.2595 ( 20.0%)  InstCombinePass
   1.4416 (  8.9%)   0.0239 ( 25.1%)   1.4655 (  9.0%)   1.4665 (  9.0%)  GVNPass
...

I perfed it and did not found any significant time eaters, all SLP-related calls took less than 2% of overall time, so no idea where to improve it for now.

[aeubanks]

I realized my timing report was slightly misleading because I had captured the IR right before SLP vectorization, so most previous passes wouldn't do any work.

If there's nothing obvious to fix then I'll close this.

[aeubanks]

Turns out it was #112221 that drastically improved compile times here.