Revive #27659: Revert "Revert "enable the SLP Vectorizer optimization pass by default""

[jrevels]

Reviving #27659 because I'd really like to see it in 0.7 if possible.

I redid the measurements from the original PR (#26594):

SLP Disabled

sysimg time:

Sysimage built. Summary:
Total ───────  99.166878 seconds
Base: ───────  26.047203 seconds 26.266%
Stdlibs: ────  73.118250 seconds 73.7325%

StaticArrays test times (I just manually put @time in front of all the relevant includes in StaticArrays' test/runtests.jl file):

@testinf      |    3      3
  1.909640 seconds (6.44 M allocations: 317.704 MiB, 4.13% gc time)
SVector       |   53     53
  1.181462 seconds (2.14 M allocations: 112.380 MiB, 1.80% gc time)
MVector       |   52     52
  0.671240 seconds (656.04 k allocations: 35.176 MiB, 1.41% gc time)
SMatrix       |   63     63
  1.455128 seconds (2.18 M allocations: 109.772 MiB, 1.69% gc time)
MMatrix       |   66     66
  1.115946 seconds (1.45 M allocations: 72.580 MiB, 1.19% gc time)
SArray        |   87     87
  4.106915 seconds (8.05 M allocations: 397.618 MiB, 3.13% gc time)
MArray        |  100    100
  3.734096 seconds (5.51 M allocations: 267.131 MiB, 1.21% gc time)
FieldVector   |   29     29
  0.812507 seconds (1.48 M allocations: 74.886 MiB, 2.14% gc time)
Scalar        |    9      9
  0.520652 seconds (1.70 M allocations: 86.460 MiB, 3.63% gc time)
SUnitRange    |   10     10
  0.096026 seconds (57.95 k allocations: 3.075 MiB)
SizedArray    |   50       1     51
  1.628736 seconds (2.69 M allocations: 135.817 MiB, 1.64% gc time)
SDiagonal     |   74     74
  9.101920 seconds (32.24 M allocations: 1.499 GiB, 5.96% gc time)
Custom types  |    2      2
  0.047831 seconds (44.94 k allocations: 2.494 MiB)
Core definitions and constructors |   79       2     81
  1.447924 seconds (995.12 k allocations: 52.970 MiB, 1.25% gc time)
AbstractArray interface |   50     50
  1.454582 seconds (1.97 M allocations: 96.524 MiB, 1.74% gc time)
Indexing      |   72     72
  2.989505 seconds (4.61 M allocations: 232.423 MiB, 2.41% gc time)
Map, reduce, mapreduce, broadcast |   67     67
  6.372919 seconds (19.25 M allocations: 956.073 MiB, 4.71% gc time)
Array math    |  141    141
  2.316072 seconds (6.94 M allocations: 353.018 MiB, 3.91% gc time)
Broadcast     |   83      12     95
  6.504455 seconds (13.22 M allocations: 691.648 MiB, 3.01% gc time)
Linear algebra |  159    159
 12.165429 seconds (28.08 M allocations: 1.363 GiB, 4.20% gc time)
Matrix multiplication |   66     66
 17.774278 seconds (38.39 M allocations: 1.783 GiB, 4.73% gc time)

SLP Enabled

sysimg time:

Sysimage built. Summary:
Total ───────  89.579786 seconds
Base: ───────  23.448816 seconds 26.1765%
Stdlibs: ────  66.129673 seconds 73.8221%

StaticArrays test times:

@testinf      |    3      3
  1.812174 seconds (6.44 M allocations: 317.704 MiB, 4.21% gc time)
SVector       |   53     53
  1.187204 seconds (2.14 M allocations: 112.379 MiB, 1.88% gc time)
MVector       |   52     52
  0.652456 seconds (656.04 k allocations: 35.176 MiB, 1.45% gc time)
SMatrix       |   63     63
  1.275547 seconds (2.18 M allocations: 109.772 MiB, 1.87% gc time)
MMatrix       |   66     66
  1.118905 seconds (1.45 M allocations: 72.611 MiB, 1.19% gc time)
SArray        |   87     87
  4.486199 seconds (8.05 M allocations: 397.271 MiB, 2.88% gc time)
MArray        |  100    100
  4.108803 seconds (5.51 M allocations: 267.131 MiB, 1.15% gc time)
FieldVector   |   29     29
  0.859377 seconds (1.48 M allocations: 74.886 MiB, 2.12% gc time)
Scalar        |    9      9
  0.562872 seconds (1.70 M allocations: 86.463 MiB, 3.37% gc time)
SUnitRange    |   10     10
  0.100489 seconds (57.95 k allocations: 3.075 MiB)
SizedArray    |   50       1     51
  1.745068 seconds (2.69 M allocations: 135.817 MiB, 1.62% gc time)
SDiagonal     |   74     74
  9.985475 seconds (32.24 M allocations: 1.499 GiB, 6.32% gc time)
Custom types  |    2      2
  0.055124 seconds (44.94 k allocations: 2.494 MiB)
Core definitions and constructors |   79       2     81
  1.575715 seconds (995.11 k allocations: 52.969 MiB, 1.55% gc time)
AbstractArray interface |   50     50
  1.553793 seconds (1.97 M allocations: 96.518 MiB, 1.98% gc time)
Indexing      |   72     72
  3.239480 seconds (4.61 M allocations: 232.406 MiB, 2.56% gc time)
Map, reduce, mapreduce, broadcast |   67     67
  7.154500 seconds (19.25 M allocations: 956.065 MiB, 5.53% gc time)
Array math    |  141    141
  2.557693 seconds (6.94 M allocations: 352.989 MiB, 3.65% gc time)
Broadcast     |   83      12     95
  7.150276 seconds (13.22 M allocations: 691.824 MiB, 2.88% gc time)
Linear algebra |  159    159
 13.607684 seconds (28.08 M allocations: 1.363 GiB, 4.11% gc time)
Matrix multiplication |   66     66
 20.916517 seconds (38.39 M allocations: 1.783 GiB, 4.47% gc time)

@nanosoldier runbenchmarks(ALL, vs = ":master")

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @ararslan

[jrevels]

I'll investigate the perf regressions and report back. It's totally possible they're real; if so, do we still want to merge this?

@andreasnoack I'm seeing a LinearAlgebra-y error on windows in this PR, I think it's unrelated but just want to check with you.

[andreasnoack]

The test error looks like an unlucky draw. Let's see if #28355 can get rid of it.

[Keno]

The skipmissing ones look somewhat interesting. Most of the other ones are among the usual candidates for flakyness.

[jrevels]

I am seeing some reproducible regressions here, but they're smaller on my machine than reported by nanosoldier (max ~10% in the skipmissing case).

That still seems worth merging to me; I'll do so tomorrow morning unless there are any objections.

[Keno]

I am seeing some reproducible regressions here, but they're smaller on my machine than reported by nanosoldier (max ~10% in the skipmissing case).

It's quite possible that this varies over architectural features. What architecture did you test with?

[jrevels]

Platform Info:
  OS: macOS (x86_64-apple-darwin17.5.0)
  CPU: Intel(R) Core(TM) i7-7920HQ CPU @ 3.10GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.0 (ORCJIT, skylake)

[jrevels]

Shall I merge this? Not sure that the (potential) regressions on nanosoldier are worth holding back here, though if others want to test more they can

[Keno]

I'm ok with this, but rc1 just went, so this'll be in rc2. Hold for a few hours since we'll be branching release-0.7 shortly.

[jrevels]

Hold for a few hours since we'll be branching release-0.7 shortly.

Okay, feel free to merge whenever. Otherwise I'll just merge tomorrow.