[PATCH] D73411: [InstCombine] Process newly inserted instructions in the correct order

[Maxim Kuvyrkov via Phabricator via llvm-commits]

maxim-kuvyrkov added a comment.

Hi Nikita,

This patch regresses one of the top-3 hot functions in 444.namd (from SPEC CPU2006).  It slows down that function by 23% (from 94 seconds to 116 seconds) and increases its code size by 13% (from 2588 bytes to 2920 bytes).  This happens at -Oz on aarch64-linux-gnu.  Since this happens at -Oz, code size regression is more interesting than the performance regression.

Would you please check if _ZN20ComputeNonbondedUtil16calc_pair_energyEP9no hits a corner case or something?  And whether this can be avoided?

444.namd,[.] _ZN20ComputeNonbondedUtil26calc_pair_energy_full,100,101,1417,1422,4120,4156
444.namd,[.] _ZN20ComputeNonbondedUtil19calc_pair_fullelectEP,101,100,1232,1240,3588,3576
444.namd,[.] _ZN20ComputeNonbondedUtil16calc_pair_energyEP9no,123,113,941,1158,2588,2920

Let me know if you need help reproducing this.  I have linux perf profiles, preprocessed source and compiled assembly handy.

Thanks!


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D73411/new/

https://reviews.llvm.org/D73411