Hunting down a compilation performance regression involving type families

[Ryan Scott]

Richard,

In an effort to figure out why programs with lots of type families have
become so much slower over the last few GHC releases, I've been looking at
GHC Trac #12545 [1], which provides a very self-contained test case that
became considerably slower to compile from GHC 7.10 to 8.0. Thanks to this
test case, I found one of the sources of the slowdown: a single, very small
commit of yours [2].

What's baffling, though, is that I can't figure out why that commit makes
compilation so much slower. I tried making profiled builds of GHC before
and after that commit and attaching SCCs to the new code, but to my
surprise, their contribution to the overall compilation time was
negligible. So I'm out of ideas on what might be causing this.

The only hint I have is the output of -ddump-simpl-stats/-dshow-passes.
Before that commit, there's a point where the terms, types, and coercions
go up:

    *** Float out(FOS {Lam = Just 0, Consts = True, OverSatApps = False}):
    Result size of Float out(FOS {Lam = Just 0,
                                  Consts = True,
                                  OverSatApps = False})
      = {terms: 20,769, types: 1,157,897, coercions: 3,635,961}
    *** Simplifier:
    Result size of Simplifier iteration=1
      = {terms: 44,081, types: 2,451,155, coercions: 9,802,016}

But after that commit, the terms and types jump up considerably higher!

    *** Simplifier:
    Result size of Simplifier
      = {terms: 16,429, types: 864,761, coercions: 2,184,448}
    *** Simplifier:
    Result size of Simplifier iteration=1
      = {terms: 87,357, types: 4,366,893, coercions: 10,008,352}

Here are the relevant bits from -ddump-simpl-stats. Before that commit:

    150 PreInlineUnconditionally
    40 PostInlineUnconditionally
    220 UnfoldingDone
    110 RuleFired
    380 BetaReduction

After that commit:

    306 PreInlineUnconditionally
    160 PostInlineUnconditionally
    360 UnfoldingDone
    308 RuleFired
    1238 BetaReduction

Does you know what might be going on here?

Ryan S.
-----
[1] https://ghc.haskell.org/trac/ghc/ticket/12545
[2]
https://ghc.haskell.org/trac/ghc/changeset/1722fa106e10e63160bb2322e2ccb830fd5b9ab3/ghc
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-devs/attachments/20170531/e5b89368/attachment.html>