Ben Gamari authored Feb 03, 2020 and Marge Bot committed Apr 07, 2020

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