Fix 635: combine boolean logic #5116

dsyme · 2018-06-06T11:50:46Z

This fixes #635 to produce better code for boolean logic. It is based on a draft by @thinkbeforecoding. The effect is to remove extra branching in nested a && b || c kind of logic, e.g.

For a variation of the repro case

let simpleAndOrEq (c1: char) (c2:char)  = ((c1 = c2) && (c1 = c2)) || (c1 = c2)

The F# code is smaller and basically identical to the equivalent C# except with one branch switched

C# IL code (size 0x14)
F# code before (size 0x1b)
F# code after (size 0x14)

In earlier examples this provided a benefit for the JITted assembly code as well, at least in size, though I haven't double checked the particular example above.

To be conservative this PR ensures no code duplication occurs, it's a strict simplification. This means that other cases like isAlphaNumeric from the linked bug are also a bit smaller but not as small as the equivalent C#. This is because the F# TAST is not fully able to express the linearization of the decision logic in this sort of code without duplication sub-expressions somewhere along the way (or using "goto"):

let isAlphaNumEq (s:string) pos =
    let c = s.[pos]
    (c = '0' && c ='9') || (c ='A' && c ='Z') || (c ='a' && c = 'z')

C# fully linearizes this. I don't think we will be able to implement this except via further work at the IlxGen.fs stage, which is independent of this PR.

TIHan · 2018-06-06T17:28:00Z

There are 4 test failures:

1) Error : FSharp-Tests-Core+TypecheckTests.type check neg17
System.Exception : neg17.err neg17.bsl differ; "diff between [D:\j\w\release_ci_pa---866fd2c3\tests\fsharp\typecheck\sigs\neg17.bsl] and [D:\j\w\release_ci_pa---866fd2c3\tests\fsharp\typecheck\sigs\neg17.err]
line 2
 - neg17.fs(84,17,84,33): typecheck error FS0292: The type definitions for type 'PrivateUnionType' in the signature and implementation are not compatible because the accessibility specified in the signature is more than that specified in the implementation
 + neg17.fs(84,17,84,33): typecheck error FS0292: The Type definitions for type 'PrivateUnionType' in the signature and implementation are not compatible because the accessibility specified in the signature is more than that specified in the implementation
"; neg17.vserr neg17.bsl differ; "diff between [D:\j\w\release_ci_pa---866fd2c3\tests\fsharp\typecheck\sigs\neg17.bsl] and [D:\j\w\release_ci_pa---866fd2c3\tests\fsharp\typecheck\sigs\neg17.vserr]
line 2
 - neg17.fs(84,17,84,33): typecheck error FS0292: The type definitions for type 'PrivateUnionType' in the signature and implementation are not compatible because the accessibility specified in the signature is more than that specified in the implementation
 + neg17.fs(84,17,84,33): typecheck error FS0292: The Type definitions for type 'PrivateUnionType' in the signature and implementation are not compatible because the accessibility specified in the signature is more than that specified in the implementation
"
   at Microsoft.FSharp.Core.PrintfModule.PrintFormatToStringThenFail@1645.Invoke(String message) in D:\j\w\release_ci_pa---866fd2c3\src\fsharp\FSharp.Core\printf.fs:line 1645

2) Error : FSharp-Tests-Core+TypecheckTests.type check neg43
System.Exception : neg43.err neg43.bsl differ; "diff between [D:\j\w\release_ci_pa---866fd2c3\tests\fsharp\typecheck\sigs\neg43.bsl] and [D:\j\w\release_ci_pa---866fd2c3\tests\fsharp\typecheck\sigs\neg43.err]
line 2
 - neg43.fs(11,6,11,8): typecheck error FS0294: The type definitions for type 'DU' in the signature and implementation are not compatible because the implementation says this type may use nulls as a representation but the signature does not
 + neg43.fs(11,6,11,8): typecheck error FS0294: The Type definitions for type 'DU' in the signature and implementation are not compatible because the implementation says this type may use nulls as a representation but the signature does not
"; neg43.vserr neg43.bsl differ; "diff between [D:\j\w\release_ci_pa---866fd2c3\tests\fsharp\typecheck\sigs\neg43.bsl] and [D:\j\w\release_ci_pa---866fd2c3\tests\fsharp\typecheck\sigs\neg43.vserr]
line 2
 - neg43.fs(11,6,11,8): typecheck error FS0294: The type definitions for type 'DU' in the signature and implementation are not compatible because the implementation says this type may use nulls as a representation but the signature does not
 + neg43.fs(11,6,11,8): typecheck error FS0294: The Type definitions for type 'DU' in the signature and implementation are not compatible because the implementation says this type may use nulls as a representation but the signature does not
"
   at Microsoft.FSharp.Core.PrintfModule.PrintFormatToStringThenFail@1645.Invoke(String message) in D:\j\w\release_ci_pa---866fd2c3\src\fsharp\FSharp.Core\printf.fs:line 1645

3) Error : FSharp-Tests-Core+TypecheckTests.type check neg57
System.Exception : neg57.err neg57.bsl differ; "diff between [D:\j\w\release_ci_pa---866fd2c3\tests\fsharp\typecheck\sigs\neg57.bsl] and [D:\j\w\release_ci_pa---866fd2c3\tests\fsharp\typecheck\sigs\neg57.err]
line 2
 - neg57.fs(4,6,4,9): typecheck error FS0314: The type definitions for type 'Foo' in the signature and implementation are not compatible because the field 'offset' was present in the implementation but not in the signature. Struct types must now reveal their fields in the signature for the type, though the fields may still be labelled 'private' or 'internal'.
 + neg57.fs(4,6,4,9): typecheck error FS0314: The Type definitions for type 'Foo' in the signature and implementation are not compatible because the field 'offset' was present in the implementation but not in the signature. Struct types must now reveal their fields in the signature for the type, though the fields may still be labelled 'private' or 'internal'.
"; neg57.vserr neg57.bsl differ; "diff between [D:\j\w\release_ci_pa---866fd2c3\tests\fsharp\typecheck\sigs\neg57.bsl] and [D:\j\w\release_ci_pa---866fd2c3\tests\fsharp\typecheck\sigs\neg57.vserr]
line 2
 - neg57.fs(4,6,4,9): typecheck error FS0314: The type definitions for type 'Foo' in the signature and implementation are not compatible because the field 'offset' was present in the implementation but not in the signature. Struct types must now reveal their fields in the signature for the type, though the fields may still be labelled 'private' or 'internal'.
 + neg57.fs(4,6,4,9): typecheck error FS0314: The Type definitions for type 'Foo' in the signature and implementation are not compatible because the field 'offset' was present in the implementation but not in the signature. Struct types must now reveal their fields in the signature for the type, though the fields may still be labelled 'private' or 'internal'.
"
   at Microsoft.FSharp.Core.PrintfModule.PrintFormatToStringThenFail@1645.Invoke(String message) in D:\j\w\release_ci_pa---866fd2c3\src\fsharp\FSharp.Core\printf.fs:line 1645

4) Error : FSharp-Tests-Core+TypecheckTests.type check neg58
System.Exception : neg58.err neg58.bsl differ; "diff between [D:\j\w\release_ci_pa---866fd2c3\tests\fsharp\typecheck\sigs\neg58.bsl] and [D:\j\w\release_ci_pa---866fd2c3\tests\fsharp\typecheck\sigs\neg58.err]
line 2
 - neg58.fs(5,6,5,9): typecheck error FS0314: The type definitions for type 'Foo' in the signature and implementation are not compatible because the field 'offset' was present in the implementation but not in the signature. Struct types must now reveal their fields in the signature for the type, though the fields may still be labelled 'private' or 'internal'.
 + neg58.fs(5,6,5,9): typecheck error FS0314: The Type definitions for type 'Foo' in the signature and implementation are not compatible because the field 'offset' was present in the implementation but not in the signature. Struct types must now reveal their fields in the signature for the type, though the fields may still be labelled 'private' or 'internal'.
"; neg58.vserr neg58.bsl differ; "diff between [D:\j\w\release_ci_pa---866fd2c3\tests\fsharp\typecheck\sigs\neg58.bsl] and [D:\j\w\release_ci_pa---866fd2c3\tests\fsharp\typecheck\sigs\neg58.vserr]
line 2
 - neg58.fs(5,6,5,9): typecheck error FS0314: The type definitions for type 'Foo' in the signature and implementation are not compatible because the field 'offset' was present in the implementation but not in the signature. Struct types must now reveal their fields in the signature for the type, though the fields may still be labelled 'private' or 'internal'.
 + neg58.fs(5,6,5,9): typecheck error FS0314: The Type definitions for type 'Foo' in the signature and implementation are not compatible because the field 'offset' was present in the implementation but not in the signature. Struct types must now reveal their fields in the signature for the type, though the fields may still be labelled 'private' or 'internal'.
"
   at Microsoft.FSharp.Core.PrintfModule.PrintFormatToStringThenFail@1645.Invoke(String message) in D:\j\w\release_ci_pa---866fd2c3\src\fsharp\FSharp.Core\printf.fs:line 1645

thinkbeforecoding · 2018-06-06T17:45:31Z

Yes, the casing of the error message changed: 'the type of' vs 'the Type of'
It seems totally orthogonal to the actual change from this PR

thinkbeforecoding · 2018-06-06T17:47:36Z

src/fsharp/tast.fs


-    override x.ToString() = 


This is the change that seems to break tests.
the casing of 'type' has changed.

Thanks for spotting this.

thinkbeforecoding · 2018-06-06T19:52:46Z

I looked at the result, and the instruction layout is not exactly the same than C#, but there is no extra instruction anymore !

Great !
After benchmarking the change, there is a small improvement in perf which was expected.

thinkbeforecoding · 2018-06-06T20:22:55Z

Using the following file:
http://www.ins.cwi.nl/projects/xmark/Assets/standard.gz
I make the following:

let txt = System.IO.File.ReadAllText "standard"
let mutable c = false
let len = txt.Length - 1
let sw = System.Diagnostics.Stopwatch.StartNew()
for _ in 1 .. 5 do
    for i in 0 .. len do
        c <- Test.isAlphaNum txt.[i]
sw.Stop()
printfn "%O" sw.Elapsed

And I get around 2.2 sec for previous version, and 2.7 sec for 'optimized' version, and 2.0 sec for cs version…
😢

vasily-kirichenko · 2018-06-06T20:30:48Z

@thinkbeforecoding Could you use BenchmarkDotNet to benchmark you sample instead?

dsyme · 2018-06-06T22:04:42Z

Using benchmark.net is sensible.

A bunch of very controlled profiling tests would be great. There are numerous variables we need to be careful about: processor, 32/64-bit, net core c netfx runtime, /optimize etc. Also branch prediction doesn't make it easy to generalise results.

dsyme · 2018-06-06T22:06:00Z

@thinkbeforecoding could you post both benchmarks where you see improvements and regressions (and is the one above really a regression? I'll check too)

thinkbeforecoding · 2018-06-06T22:12:10Z

I'm working on a better benchmark with benchmarkdotnet

thinkbeforecoding · 2018-06-06T23:07:22Z

BenchmarkDotNet is giving:
Optimized FS: 771.4ms
Original FS: 608.7ms
Cs: 539.9ms

This is testing applying isAlphaNum on the given file...

dsyme · 2018-06-06T23:28:20Z

@thinkbeforecoding thanks for testing. Please link a zip and give full details of OS, .NET runtime, architecture, runtime etc. thx

dsyme · 2018-06-06T23:30:48Z

@thinkbeforecoding I couldn't unzip that archive - could you post it with ZIP please? thanks

dsyme · 2018-06-06T23:43:10Z

@thinkbeforecoding On a quick test I got this on Windows 10 .NET 4.7.2 Intel(R) Xeon(R) CPU E5-1620 0 @ 3.60GHz, 3601 Mhz, 4 Core(s), 8 Logical Processor(s)

x64:

F# VS 15.7.2: 4.0
F# THIS PR: 3.7

x32:

F# VS 15.7.2: 5.1
F# THIS PR: 4.3

Those are just approximate but the numbers are pretty consistent. That's for this code:

let isAlphaNum (s:string) pos =
    let c = s.[pos]
    (c >= '0' && c<='9') || (c>='A' && c<='Z') || (c>='a' && c <= 'z')
    
[<EntryPoint>]
let main argv = 

    let txt = System.IO.File.ReadAllText @"c:\GitHub\dsyme\NInterpret\Platforms\NInterpret.Xamarin.Droid\obj\Debug\lp\16\jl\bin\R.txt"
    let mutable c = false
    let len = txt.Length - 1
    let sw = System.Diagnostics.Stopwatch.StartNew()
    for _ in 1 .. 10000 do
        for i in 0 .. len do
            c <- isAlphaNum txt i
    sw.Stop()
    printfn "%O" sw.Elapsed

    0

Compiling and running with

debug\net40\bin\fsc /platform:x64 a.fs && a.exe
debug\net40\bin\fsc /platform:x86 a.fs && a.exe
fsc /platform:x64 a.fs && a.exe
fsc /platform:x86 a.fs && a.exe

So that looks like speed up to me, but we should continue to test

thinkbeforecoding · 2018-06-07T06:28:02Z

Ok, I try to test more cases today 👍🏻 jeremie chassaing / thinkbeforecoding

thinkbeforecoding · 2018-06-07T08:46:00Z

The .net core test on the same sample is giving similar results:

Intel Core i7-4650U CPU 1.70GHz (Haswell), 1 CPU, 4 logical and 2 physical cores
.NET Core SDK=2.1.300
  [Host]     : .NET Core 2.1.0 (CoreCLR 4.6.26515.07, CoreFX 4.6.26515.06), 64bit RyuJIT DEBUG
  DefaultJob : .NET Core 2.1.0 (CoreCLR 4.6.26515.07, CoreFX 4.6.26515.06), 64bit RyuJIT

Optimized F#: 646.9ms
C#: 543.7ms
Original F#: 599.1ms

thinkbeforecoding · 2018-06-07T10:05:56Z

More benchmark:

BenchmarkDotNet=v0.10.14, OS=Windows 10.0.17134
Intel Core i7-4650U CPU 1.70GHz (Haswell), 1 CPU, 4 logical and 2 physical cores
.NET Core SDK=2.1.300
  [Host] : .NET Core 2.1.0 (CoreCLR 4.6.26515.07, CoreFX 4.6.26515.06), 64bit RyuJIT DEBUG
  Clr    : .NET Framework 4.7.1 (CLR 4.0.30319.42000), 64bit RyuJIT-v4.7.3110.0
  Core   : .NET Core 2.1.0 (CoreCLR 4.6.26515.07, CoreFX 4.6.26515.06), 64bit RyuJIT

Method	Job	Runtime	Categories	Mean	Error	StdDev	Scaled	ScaledSD
Optimized_AlphaNum	Clr	Clr	AlphaNum	97.15 ms	1.6824 ms	1.2165 ms	0.93	0.01
Cs_AlphaNum	Clr	Clr	AlphaNum	104.29 ms	0.9284 ms	0.6713 ms	1.00	0.00
Original_AlphaNum	Clr	Clr	AlphaNum	98.98 ms	1.9299 ms	2.1451 ms	0.95	0.02

Optimized_NumAlpha	Clr	Clr	NumAlpha	121.61 ms	1.7726 ms	1.5713 ms	1.16	0.03
Cs_NumAlpha	Clr	Clr	NumAlpha	104.48 ms	2.3507 ms	2.0839 ms	1.00	0.00
Original_NumAlpha	Clr	Clr	NumAlpha	139.68 ms	3.9307 ms	9.3416 ms	1.34	0.09

Optimized_AlphaNum	Core	Core	AlphaNum	119.88 ms	2.2961 ms	2.0355 ms	0.92	0.02
Cs_AlphaNum	Core	Core	AlphaNum	130.87 ms	0.3338 ms	0.2959 ms	1.00	0.00
Original_AlphaNum	Core	Core	AlphaNum	119.79 ms	0.5994 ms	0.5313 ms	0.92	0.00

Optimized_NumAlpha	Core	Core	NumAlpha	168.34 ms	0.7185 ms	0.6369 ms	1.42	0.01
Cs_NumAlpha	Core	Core	NumAlpha	118.27 ms	1.1111 ms	0.9849 ms	1.00	0.00
Original_NumAlpha	Core	Core	NumAlpha	130.85 ms	2.5325 ms	2.1148 ms	1.11	0.02

The code:

[<CoreJob; ClrJob;GroupBenchmarksBy(BenchmarkLogicalGroupRule.ByCategory);CategoriesColumn>]
type BoolPerf2() =
    let txt = System.IO.File.ReadAllText @"visualfsharp\packages\StrawberryPerl64.5.22.2.1\Tools\c\x86_64-w64-mingw32\include\mshtml.h"
    let len = txt.Length - 1
    [<Benchmark; BenchmarkCategory("AlphaNum")>]
    member __.Optimized_AlphaNum() = 
        for _ in 1 ..3 do
        for i in 0 .. len do
            TestOptimized.isAlphaNum txt.[i] |> ignore
    [<Benchmark; BenchmarkCategory("NumAlpha")>]
    member __.Optimized_NumAlpha() = 
        for _ in 1 ..3 do
        for i in 0 .. len do
            TestOptimized.isNumAlpha txt.[i] |> ignore
    [<Benchmark(Baseline = true); BenchmarkCategory("AlphaNum")>]
    member __.Cs_AlphaNum() = 
        for _ in 1 ..3 do
        for i in 0 .. len do
            TestCs.isAlphaNum txt.[i] |> ignore
    [<Benchmark(Baseline = true); BenchmarkCategory("NumAlpha")>]
    member __.Cs_NumAlpha() = 
        for _ in 1 ..3 do
        for i in 0 .. len do
            TestCs.isNumAlpha txt.[i] |> ignore
    [<Benchmark; BenchmarkCategory("AlphaNum")>]
    member __.Original_AlphaNum() = 
        for _ in 1 ..3 do
        for i in 0 .. len do
            Test.isAlphaNum txt.[i] |> ignore
    [<Benchmark; BenchmarkCategory("NumAlpha")>]
    member __.Original_NumAlpha() = 
        for _ in 1 ..3 do
        for i in 0 .. len do
            Test.isNumAlpha txt.[i] |> ignore


And the implementation of isAlphaNum and isNumAlpha:

``` F#
let isAlphaNum c =
    (c>='a' && c<='z') || (c >= '0' && c<='9')
let isNumAlpha c =
    (c >= '0' && c<='9') || (c>='a' && c<='z')


So the optimized version is sometime better than C# version. Depends highly on final code layout and CPU branch prediction based on input data.

…o bool2

thinkbeforecoding · 2018-06-07T19:30:57Z

Great !

KevinRansom · 2018-09-12T20:59:47Z

@dsyme, is this ready to go?

dsyme · 2018-09-13T13:12:04Z

Note to self: baselines need updating:

2018-09-13T09:18:26.3121234Z CodeGen\EmittedIL\Mutation (Mutation02.fs) -- failed
2018-09-13T09:18:26.3121812Z CodeGen\EmittedIL\Mutation (Mutation03.fs) -- failed
2018-09-13T09:18:26.3122366Z CodeGen\EmittedIL\Mutation (Mutation04.fs) -- failed
2018-09-13T09:18:26.3122948Z CodeGen\EmittedIL\Mutation (Mutation05.fs) -- failed

dsyme · 2018-09-13T14:36:23Z

OK, this is ready to go

thinkbeforecoding · 2018-09-13T14:56:02Z

Yeah, cool ! Thx jeremie chassaing / thinkbeforecoding De : Don Syme Envoyé le :jeudi 13 septembre 2018 16:37 À : Microsoft/visualfsharp Cc : Jérémie Chassaing; Mention Objet :Re: [Microsoft/visualfsharp] Fix 635: combine boolean logic (#5116) OK, this is ready to go — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

KevinRansom · 2018-09-13T20:17:57Z

@dsyme …. misery …. your last pr that I just merged stepped over these baselines.

KevinRansom · 2018-09-13T20:26:43Z

@dsyme, I hope I resolved them right.

Thorium · 2018-09-16T12:41:35Z

Maybe off-topic but the title sounds bit like what I did for LINQ-expression trees (and the same could be done for F# Exprs): https://thorium.github.io/Linq.Expression.Optimizer/

dsyme added 2 commits June 6, 2018 12:33

combine boolean logic

4fdc34b

generalize and ensure no code duplication

d12b524

dsyme changed the title ~~Fix 635: combine boolean logic~~ [CompilerPerf] Fix 635: combine boolean logic Jun 6, 2018

thinkbeforecoding reviewed Jun 6, 2018

View reviewed changes

fix tests

00e6d62

dsyme added 3 commits June 7, 2018 11:59

Merge branch 'master' of http://github.com/Microsoft/visualfsharp int…

39bad3f

…o bool2

update baselines

18e4e2a

fix baseline

74c98a8

dsyme added 2 commits June 7, 2018 21:00

Update Optimizer.fs

1212562

Merge branch 'master' into bool2

6eaedcf

dsyme changed the title ~~[CompilerPerf] Fix 635: combine boolean logic~~ Fix 635: combine boolean logic Jun 8, 2018

cartermp added the Tenet-Performance label Jul 21, 2018

cartermp added this to the 16.0 milestone Jul 21, 2018

KevinRansom closed this Sep 6, 2018

KevinRansom reopened this Sep 6, 2018

KevinRansom closed this Sep 12, 2018

KevinRansom reopened this Sep 12, 2018

KevinRansom approved these changes Sep 12, 2018

View reviewed changes

merge master

afd9ef4

update baselines

2b80cf0

Merge branch 'master' into bool2

c3ade2e

KevinRansom merged commit 5e8352b into dotnet:master Sep 15, 2018

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fix 635: combine boolean logic #5116

Fix 635: combine boolean logic #5116

dsyme commented Jun 6, 2018 •

edited

Loading

TIHan commented Jun 6, 2018

thinkbeforecoding commented Jun 6, 2018

thinkbeforecoding Jun 6, 2018

dsyme Jun 6, 2018

thinkbeforecoding commented Jun 6, 2018

thinkbeforecoding commented Jun 6, 2018

vasily-kirichenko commented Jun 6, 2018

dsyme commented Jun 6, 2018

dsyme commented Jun 6, 2018

thinkbeforecoding commented Jun 6, 2018

thinkbeforecoding commented Jun 6, 2018

dsyme commented Jun 6, 2018

dsyme commented Jun 6, 2018

dsyme commented Jun 6, 2018 •

edited

Loading

thinkbeforecoding commented Jun 7, 2018 via email •

edited by dsyme

Loading

thinkbeforecoding commented Jun 7, 2018 •

edited

Loading

thinkbeforecoding commented Jun 7, 2018 •

edited

Loading

thinkbeforecoding commented Jun 7, 2018

KevinRansom commented Sep 12, 2018

dsyme commented Sep 13, 2018

dsyme commented Sep 13, 2018

thinkbeforecoding commented Sep 13, 2018 via email

KevinRansom commented Sep 13, 2018

KevinRansom commented Sep 13, 2018

Thorium commented Sep 16, 2018

Fix 635: combine boolean logic #5116

Fix 635: combine boolean logic #5116

Conversation

dsyme commented Jun 6, 2018 • edited Loading

TIHan commented Jun 6, 2018

thinkbeforecoding commented Jun 6, 2018

thinkbeforecoding Jun 6, 2018

Choose a reason for hiding this comment

dsyme Jun 6, 2018

Choose a reason for hiding this comment

thinkbeforecoding commented Jun 6, 2018

thinkbeforecoding commented Jun 6, 2018

vasily-kirichenko commented Jun 6, 2018

dsyme commented Jun 6, 2018

dsyme commented Jun 6, 2018

thinkbeforecoding commented Jun 6, 2018

thinkbeforecoding commented Jun 6, 2018

dsyme commented Jun 6, 2018

dsyme commented Jun 6, 2018

dsyme commented Jun 6, 2018 • edited Loading

thinkbeforecoding commented Jun 7, 2018 via email • edited by dsyme Loading

thinkbeforecoding commented Jun 7, 2018 • edited Loading

thinkbeforecoding commented Jun 7, 2018 • edited Loading

thinkbeforecoding commented Jun 7, 2018

KevinRansom commented Sep 12, 2018

dsyme commented Sep 13, 2018

dsyme commented Sep 13, 2018

thinkbeforecoding commented Sep 13, 2018 via email

KevinRansom commented Sep 13, 2018

KevinRansom commented Sep 13, 2018

Thorium commented Sep 16, 2018

dsyme commented Jun 6, 2018 •

edited

Loading

dsyme commented Jun 6, 2018 •

edited

Loading

thinkbeforecoding commented Jun 7, 2018 via email •

edited by dsyme

Loading

thinkbeforecoding commented Jun 7, 2018 •

edited

Loading

thinkbeforecoding commented Jun 7, 2018 •

edited

Loading