CmmPipeline: add a second pass of CmmCommonBlockElim
authorMichal Terepeta <michal.terepeta@gmail.com>
Mon, 26 Mar 2018 19:18:44 +0000 (15:18 -0400)
committerBen Gamari <ben@smart-cactus.org>
Tue, 27 Mar 2018 13:53:10 +0000 (09:53 -0400)
commitd5c4d46a62ce6a0cfa6440344f707136eff18119
tree539f6eaf8fa90d791ab8315b6f40c8687f56e92b
parent8cfd2e4e5832bd188816e64daaa95c9003437723
CmmPipeline: add a second pass of CmmCommonBlockElim

The sinking pass often gets rid of unnecessary registers
registers/assignements exposing more opportunities for CBE, so this
commit adds a second round of CBE after the sinking pass and should
fix #12915 (and some examples in #14226).

Nofib results:
* Binary size:         0.9% reduction on average
* Compile allocations: 0.7% increase on average
* Runtime:             noisy, two separate runs of nofib showed a tiny
                       reduction on average, (~0.2-0.3%), but I think
                       this is mostly noise
* Compile time:        very noisy, but generally within +/- 0.5% (one
                       run faster, one slower)

One interesting part of this change is that running CBE invalidates
results of proc-point analysis. But instead of re-doing the whole
analysis, we can use the map that CBE creates for replacing/comparing
block labels (maps a redundant label to a useful one) to update the
results of proc-point analysis. This lowers the overhead compared to the
previous experiment in #12915.

Signed-off-by: Michal Terepeta <michal.terepeta@gmail.com>
Test Plan: ./validate

Reviewers: bgamari, simonmar

Reviewed By: simonmar

Subscribers: rwbarton, thomie, carter

GHC Trac Issues: #12915, #14226

Differential Revision: https://phabricator.haskell.org/D4417
compiler/cmm/CmmCommonBlockElim.hs
compiler/cmm/CmmPipeline.hs