<!DOCTYPE html PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN" "http://www.w3.org/TR/REC-html40/loose.dtd">
<html lang="en">
<head>
<meta content="text/html; charset=US-ASCII" http-equiv="Content-Type">
<title>
GitLab
</title>



<style>img {
max-width: 100%; height: auto;
}
</style>
</head>
<body>
<div class="content">

<h3>
Ryan Scott pushed to branch wip/T18023
at <a href="https://gitlab.haskell.org/ghc/ghc">Glasgow Haskell Compiler / GHC</a>
</h3>
<h4>
Commits:
</h4>
<ul>
<li>
<strong><a href="https://gitlab.haskell.org/ghc/ghc/-/commit/f38e8d61f066c3064c600c352eebcd87f28d989a">f38e8d61</a></strong>
<div>
<span>by Daniel Gröber</span>
<i>at 2020-04-07T02:00:05-04:00</i>
</div>
<pre class="commit-message" style="white-space: pre-wrap; margin: 0;">rts: ProfHeap: Fix memory leak when not compiled with profiling

If we're doing heap profiling on an unprofiled executable we keep
allocating new space in initEra via nextEra on each profiler run but we
don't have a corresponding freeEra call.

We do free the last era in endHeapProfiling but previous eras will have
been overwritten by initEra and will never get free()ed.

Metric Decrease:
    space_leak_001
</pre>
</li>
<li>
<strong><a href="https://gitlab.haskell.org/ghc/ghc/-/commit/bcd668593e1c9e00c5d9c72960b4833dd526cb9a">bcd66859</a></strong>
<div>
<span>by Sebastian Graf</span>
<i>at 2020-04-07T02:00:41-04:00</i>
</div>
<pre class="commit-message" style="white-space: pre-wrap; margin: 0;">Re-export GHC.Magic.noinline from base
</pre>
</li>
<li>
<strong><a href="https://gitlab.haskell.org/ghc/ghc/-/commit/3d2991f8b4c1b686323b2c9452ce845a60b8d94c">3d2991f8</a></strong>
<div>
<span>by Ben Gamari</span>
<i>at 2020-04-07T18:36:09-04:00</i>
</div>
<pre class="commit-message" style="white-space: pre-wrap; margin: 0;">simplifier: Kill off ufKeenessFactor

We used to have another factor, ufKeenessFactor, which would scale the
discounts before they were subtracted from the size. This was justified
with the following comment:

  -- We multiple the raw discounts (args_discount and result_discount)
  -- ty opt_UnfoldingKeenessFactor because the former have to do with
  --  *size* whereas the discounts imply that there's some extra
  --  *efficiency* to be gained (e.g. beta reductions, case reductions)
  -- by inlining.

However, this is highly suspect since it means that we subtract a
*scaled* size from an absolute size, resulting in crazy (e.g. negative)
scores in some cases (#15304). We consequently killed off
ufKeenessFactor and bumped up the ufUseThreshold to compensate.

Adjustment of unfolding use threshold
=====================================

Since this removes a discount from our inlining heuristic, I revisited our
default choice of -funfolding-use-threshold to minimize the change in
overall inlining behavior. Specifically, I measured runtime allocations
and executable size of nofib and the testsuite performance tests built
using compilers (and core libraries) built with several values of
-funfolding-use-threshold.

This comes as a result of a quantitative comparison of testsuite
performance and code size as a function of ufUseThreshold, comparing
GHC trees using values of 50, 60, 70, 80, 90, and 100. The test set
consisted of nofib and the testsuite performance tests.
A full summary of these measurements are found in the description of
!2608

Comparing executable sizes (relative to the base commit) across all
nofib tests, we see that sizes are similar to the baseline:

            gmean      min      max   median
thresh
50         -6.36%   -7.04%   -4.82%   -6.46%
60         -5.04%   -5.97%   -3.83%   -5.11%
70         -2.90%   -3.84%   -2.31%   -2.92%
80         -0.75%   -2.16%   -0.42%   -0.73%
90         +0.24%   -0.41%   +0.55%   +0.26%
100        +1.36%   +0.80%   +1.64%   +1.37%
baseline   +0.00%   +0.00%   +0.00%   +0.00%

Likewise, looking at runtime allocations we see that 80 gives slightly
better optimisation than the baseline:

            gmean      min      max   median
thresh
50         +0.16%   -0.16%   +4.43%   +0.00%
60         +0.09%   -0.00%   +3.10%   +0.00%
70         +0.04%   -0.09%   +2.29%   +0.00%
80         +0.02%   -1.17%   +2.29%   +0.00%
90         -0.02%   -2.59%   +1.86%   +0.00%
100        +0.00%   -2.59%   +7.51%   -0.00%
baseline   +0.00%   +0.00%   +0.00%   +0.00%

Finally, I had to add a NOINLINE in T4306 to ensure that `upd` is
worker-wrappered as the test expects. This makes me wonder whether the
inlining heuristic is now too liberal as `upd` is quite a large
function. The same measure was taken in T12600.

             Wall clock time compiling Cabal with -O0
thresh       50     60     70     80     90      100    baseline
build-Cabal  93.88  89.58  92.59  90.09  100.26  94.81  89.13

Also, this change happens to avoid the spurious test output in
`plugin-recomp-change` and `plugin-recomp-change-prof` (see #17308).

Metric Decrease:
    hie002
    T12234
    T13035
    T13719
    T14683
    T4801
    T5631
    T5642
    T9020
    T9872d
    T9961
Metric Increase:
    T12150
    T12425
    T13701
    T14697
    T15426
    T1969
    T3064
    T5837
    T6048
    T9203
    T9872a
    T9872b
    T9872c
    T9872d
    haddock.Cabal
    haddock.base
    haddock.compiler
</pre>
</li>
<li>
<strong><a href="https://gitlab.haskell.org/ghc/ghc/-/commit/255418da5d264fb2758bc70925adb2094f34adc3">255418da</a></strong>
<div>
<span>by Sylvain Henry</span>
<i>at 2020-04-07T18:36:49-04:00</i>
</div>
<pre class="commit-message" style="white-space: pre-wrap; margin: 0;">Modules: type-checker (#13009)

Update Haddock submodule
</pre>
</li>
<li>
<strong><a href="https://gitlab.haskell.org/ghc/ghc/-/commit/6c697c24cda33fb7ec5b9ff36188584ea8d3e81a">6c697c24</a></strong>
<div>
<span>by Ryan Scott</span>
<i>at 2020-04-07T19:45:59-04:00</i>
</div>
<pre class="commit-message" style="white-space: pre-wrap; margin: 0;">Use conLikeUserTyVarBinders to quantify field selector types

This patch:

1. Writes up a specification for how the types of top-level field
   selectors should be determined in a new section of the GHC User's
   Guide, and
2. Makes GHC actually implement that specification by using
   `conLikeUserTyVarBinders` in `mkOneRecordSelector` to preserve the
   order and specificity of type variables written by the user.

Fixes #18023.
</pre>
</li>
</ul>
<h4>30 changed files:</h4>
<ul>
<li class="file-stats">
<a href="#abe895765c4ce08c5dfbc95e5e3c5db67ff92ded">
CODEOWNERS
</a>
</li>
<li class="file-stats">
<a href="#d0d96a6d03668aeab20ebe05e2c4ccb798c7e64c">
compiler/GHC.hs
</a>
</li>
<li class="file-stats">
<a href="#56e23d78cfece2c83f03ed9b9a8ce9b20be26462">
compiler/GHC/Cmm/Expr.hs
</a>
</li>
<li class="file-stats">
<a href="#90378e83c3a00a78bc0b3c01da111e0a787de451">
compiler/GHC/Cmm/Node.hs
</a>
</li>
<li class="file-stats">
<a href="#182d6a315e784018aa9c8b2ad736036b97bd5d48">
compiler/GHC/Core.hs
</a>
</li>
<li class="file-stats">
<a href="#8b1482a65a3092c268b853a00aa18286427d1dec">
compiler/GHC/Core/Arity.hs
</a>
</li>
<li class="file-stats">
<a href="#e39f2416dd3b2af571ddd26dcca3476a55810c41">
compiler/GHC/Core/Class.hs
</a>
</li>
<li class="file-stats">
<a href="#783e5dae6e86931f06700fc088fb7d48c8a07386">
compiler/GHC/Core/Coercion.hs
</a>
</li>
<li class="file-stats">
<a href="#f10ed7a2470454dfdd8691a08beba67d8b78ee70">
compiler/GHC/Core/Coercion/Axiom.hs
</a>
</li>
<li class="file-stats">
<a href="#975dc08a8e7942b32d621f617d5a9c1b668601dd">
compiler/GHC/Core/Coercion/Opt.hs
</a>
</li>
<li class="file-stats">
<a href="#b452cc5181a4af3460c43a4e3e4b5edaa238b4e0">
compiler/GHC/Core/ConLike.hs
</a>
</li>
<li class="file-stats">
<a href="#6fcf64907fb5bdd93082d2d1eb94e4566e735865">
compiler/GHC/Core/DataCon.hs
</a>
</li>
<li class="file-stats">
<a href="#91648438362e5a35363d2bb7abb04016dedd7d7e">
compiler/GHC/Core/FamInstEnv.hs
</a>
</li>
<li class="file-stats">
<a href="#5596bb0f228bd2a308c4391df97375ae879430a8">
compiler/GHC/Core/InstEnv.hs
</a>
</li>
<li class="file-stats">
<a href="#36a42448a83a9d1f6df8475f03ead2eed199dd8e">
compiler/GHC/Core/Lint.hs
</a>
</li>
<li class="file-stats">
<a href="#e5041b26fef1c846489ee5e1c883bc3d52cc775b">
compiler/GHC/Core/Op/CSE.hs
</a>
</li>
<li class="file-stats">
<a href="#9e2ce77a946950b2d19d75c55f671c048a5645bd">
compiler/GHC/Core/Op/OccurAnal.hs
</a>
</li>
<li class="file-stats">
<a href="#e3fe84599ec044fc7aacb062dada32b067732fc8">
compiler/GHC/Core/Op/Simplify.hs
</a>
</li>
<li class="file-stats">
<a href="#d52bc1e90d7432064a2998de2a2751b6fecb7680">
compiler/GHC/Core/Op/Specialise.hs
</a>
</li>
<li class="file-stats">
<a href="#5e5f8dbd247535be3f94c4bac0005587d000dba0">
compiler/GHC/Core/PatSyn.hs
</a>
</li>
<li class="file-stats">
<a href="#0ca1508e5b0cedec2f19fe10e414597184209573">
compiler/GHC/Core/Predicate.hs
</a>
</li>
<li class="file-stats">
<a href="#d01de95d8c71d4dec6a9a59146186e29a5b3fd38">
compiler/GHC/Core/Rules.hs
</a>
</li>
<li class="file-stats">
<a href="#039e8f5676356ef05c90af828aae48aac7296e47">
compiler/GHC/Core/TyCo/FVs.hs
</a>
</li>
<li class="file-stats">
<a href="#379343fa560df0c7451aa5a1dd4f074dd37924d0">
compiler/GHC/Core/TyCo/Ppr.hs
</a>
</li>
<li class="file-stats">
<a href="#578c7c3857d66d963736ff6742f3433a0e8d01b7">
compiler/GHC/Core/TyCo/Rep.hs
</a>
</li>
<li class="file-stats">
<a href="#4aad0050db1a8a20db8bbca149111de99cb299c9">
compiler/GHC/Core/TyCon.hs
</a>
</li>
<li class="file-stats">
<a href="#84c9d3ba34173297e7d7747e02caa2e004b164fc">
compiler/GHC/Core/Type.hs
</a>
</li>
<li class="file-stats">
<a href="#2811a7297b8aa206197ac1f5dabd0818e3c7ec5a">
compiler/GHC/Core/Unfold.hs
</a>
</li>
<li class="file-stats">
<a href="#89a2e0e1a3095c7a42b10f2819049908908080e8">
compiler/GHC/Core/Unify.hs
</a>
</li>
<li class="file-stats">
<a href="#a5d27de4e3111a69f69902c65ecf7b3379f7c49a">
compiler/GHC/CoreToStg/Prep.hs
</a>
</li>
</ul>
<h5>The diff was not included because it is too large.</h5>

</div>
<div class="footer" style="margin-top: 10px;">
<p style="font-size: small; color: #777;">

<br>
<a href="https://gitlab.haskell.org/ghc/ghc/-/compare/d3706e35310d150f6ecb4eeade0c3e1b7108d9cf...6c697c24cda33fb7ec5b9ff36188584ea8d3e81a">View it on GitLab</a>.
<br>
You're receiving this email because of your account on gitlab.haskell.org.
If you'd like to receive fewer emails, you can
adjust your notification settings.



</p>
</div>
</body>
</html>