Skip to content
Snippets Groups Projects
Commit 96d32ff2 authored by Simon Peyton Jones's avatar Simon Peyton Jones Committed by Simon Peyton Jones
Browse files

Make rewrite rules "win" over inlining

If a rewrite rule and a rewrite rule compete in the simplifier, this
patch makes sure that the rewrite rule "win".  That is, in general
a bit fragile, but it's a huge help when making specialisation work
reliably, as #21851 and #22097 showed.

The change is fairly straightforwad, and documented in
   Note [Rewrite rules and inlining]
in GHC.Core.Opt.Simplify.Iteration.

Compile-times change, up and down a bit -- in some cases because
we get better specialisation.  But the payoff (more reliable
specialisation) is large.

Metrics: compile_time/bytes allocated
-----------------------------------------------
    T10421(normal)   +3.7% BAD
   T10421a(normal)   +5.5%
    T13253(normal)   +1.3%
      T14052(ghci)   +1.8%
    T15304(normal)   -1.4%
    T16577(normal)   +3.1% BAD
    T17516(normal)   +2.3%
    T17836(normal)   -1.9%
    T18223(normal)   -1.8%
     T8095(normal)   -1.3%
     T9961(normal)   +2.5% BAD

         geo. mean   +0.0%
         minimum     -1.9%
         maximum     +5.5%

Nofib results are (bytes allocated)

+-------------------------------++----------+
|                               ||tsv (rel) |
+===============================++==========+
|           imaginary/paraffins ||   +0.27% |
|                imaginary/rfib ||   -0.04% |
|                     real/anna ||   +0.02% |
|                      real/fem ||   -0.04% |
|                    real/fluid ||   +1.68% |
|                   real/gamteb ||   -0.34% |
|                       real/gg ||   +1.54% |
|                   real/hidden ||   -0.01% |
|                      real/hpg ||   -0.03% |
|                    real/infer ||   -0.03% |
|                   real/prolog ||   +0.02% |
|                  real/veritas ||   -0.47% |
|       shootout/fannkuch-redux ||   -0.03% |
|         shootout/k-nucleotide ||   -0.02% |
|               shootout/n-body ||   -0.06% |
|        shootout/spectral-norm ||   -0.01% |
|         spectral/cryptarithm2 ||   +1.25% |
|             spectral/fibheaps ||  +18.33% |
|           spectral/last-piece ||   -0.34% |
+===============================++==========+
|                     geom mean ||   +0.17% |

There are extensive notes in !8897 about the regressions.
Briefly

* fibheaps: there was a very delicately balanced inlining that
  tipped over the wrong way after this change.

* cryptarithm2 and paraffins are caused by #22274, which is
  a separate issue really.  (I.e. the right fix is *not* to
  make inlining "win" over rules.)

So I'm accepting these changes

Metric Increase:
    T10421
    T16577
    T9961
parent 44fcdb04
No related branches found
No related tags found
No related merge requests found
Showing
with 344 additions and 161 deletions
Loading
0% Loading or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment