Addresses Regex Perf Issue 32764 #32899

Alois-xx · 2018-10-18T21:13:10Z

This fix should save ca 40% for the common case when many Char.ToLower calls are emitted which access CultureInfo.CurrentCulture for compiled Regex Queries.

…xOptions.IgnoreCase or RegexOptions.CultureInvariant. This saves over 40% in these cases.

……xOptions.IgnoreCase or RegexOptions.CultureInvariant. This saves over 40% in these cases.

…ith RegexOptions.IgnoreCase or RegexOptions.CultureInvariant." This reverts commit e151e93.

stephentoub · 2018-10-18T21:33:45Z

src/System.Text.RegularExpressions/src/System/Text/RegularExpressions/RegexCompiler.cs

+
+        private void CallToLower()
+        {
+             Ldloc(_cultureV);


What is the lifetime of this? e.g. does this cache end up spanning multiple regex calls and user code such that user code could change the current culture and end up with different behavior than before?

_cultureV is a local variable. The lifetime of the cache is therefore method local in FindFirstChar and Go methods of the compiled regular expression.

Ah, I see. Thanks.

stephentoub · 2018-10-18T21:35:02Z

src/System.Text.RegularExpressions/src/System/Text/RegularExpressions/RegexCompiler.cs

+        private void CallToLower()
+        {
+             Ldloc(_cultureV);
+             Call(s_chartolowerM);


@ViktorHofer, am I misremembering, or did you do some ToLower-related optimizations/changes when you were cleaning up the regex code recently? If yes, should we instead just port the appropriate changes to the compiled version? If no, is there a similar optimization to be done on the interpreted side of things?

Right, but only minor allocation optimizations. I wouldn't do that as part of this PR but revisit the compiled code paths later to make sure that we don't diverge and bring optimizations over.

stephentoub · 2018-10-19T11:09:38Z

Do we have regex tests (interpreted and compiled) that rely on the current culture being something specific? If not, it'd be good to add as part of this.

ViktorHofer · 2018-10-19T15:30:22Z

Do we have regex tests (interpreted and compiled) that rely on the current culture being something specific? If not, it'd be good to add as part of this.

Only very few for Unicode characters. We should definitely add a bunch.

ViktorHofer

Can you please post performance numbers with BenchmarkDotNet baselined on master including allocation?

Alois-xx · 2018-10-19T22:34:02Z

@ViktorHofer
Benchmark.NET was a little funny to get it to use the compiled dlls and not the ones from the framework sdk. Here are the patched Results:

                        Method | N |         Mean |        Error |       StdDev | Allocated |
------------------------------ |-- |-------------:|-------------:|-------------:|----------:|
                      Original | 1 | 438,212.4 ns | 4,647.930 ns | 4,347.676 ns |       0 B |
                 CaseSensitive | 1 | 227,563.0 ns | 5,142.693 ns | 5,281.167 ns |       0 B |
 CaseSensitiveCultureInvariant | 1 | 215,541.1 ns | 1,187.442 ns | 1,110.734 ns |       0 B |
          SimplifiedExpression | 1 |   1,525.5 ns |    11.969 ns |    11.195 ns |       0 B |
      Case_Sensitive_Substring | 1 |     691.5 ns |     4.897 ns |     4.581 ns |       0 B |

And here the baseline version where I did revert the changes of my commit to really measure the impact of only my change:

                        Method | N |           Mean |         Error |        StdDev | Allocated |
------------------------------ |-- |---------------:|--------------:|--------------:|----------:|
                      Original | 1 | 1,055,537.2 ns | 31,271.169 ns | 36,011.922 ns |       0 B |
                 CaseSensitive | 1 |   216,579.6 ns |  1,827.153 ns |  1,709.119 ns |       0 B |
 CaseSensitiveCultureInvariant | 1 |   217,116.7 ns |  2,212.376 ns |  1,961.213 ns |       0 B |
          SimplifiedExpression | 1 |     1,523.0 ns |     13.191 ns |     12.339 ns |       0 B |
      Case_Sensitive_Substring | 1 |       671.3 ns |      8.552 ns |      8.000 ns |       0 B |

Just in case you did not see here is the ETW chart:

Actually the Benchmark.NET numbers are even better. From the numbers it looks like the slow code gen path is only triggerd when RegexOptions.IgnoreCase is used. But anyway this is certainly a widely used option.

Alois-xx · 2018-10-20T15:06:42Z

@stephentoub: I could think of some Turkish I tests which behave differently in different locales. That would be a good test to verify that the right culture was used.

ViktorHofer · 2018-10-22T16:00:53Z

That sounds good and should suffice. As mentioned before, we currently don't have a comprehensive set of inputs with different cultures. We should definitely fix that.

…ressions.

Alois-xx · 2018-10-22T21:55:56Z

@ViktorHofer: Added some locale tests which really show different behavior under the turkish locale.

Alois-xx · 2018-10-24T19:38:47Z

@dotnet-bot test this please

stephentoub · 2018-10-25T05:14:47Z

src/System.Text.RegularExpressions/tests/RegexCultureTests.cs

+            string input = "Iıİi";
+
+            var cultInvariantRegex = Create(input, CultureInfo.InvariantCulture, RegexOptions.IgnoreCase | RegexOptions.CultureInvariant);
+            var turkishRegex =       Create(input, turkish,                      RegexOptions.IgnoreCase);


Nit: please change var to be the actual type here... we only use var when the type is obvious from the right-hand side, namely a ctor or an explicit cast.

Alois-xx · 2018-10-25T11:53:35Z

@dotnet-bot test this please

stephentoub · 2018-10-25T14:54:42Z

Thanks!

danmoseley · 2018-10-25T15:44:14Z

Nice win here @Alois-xx many thanks! Any interest in more work on regex?

Alois-xx · 2018-10-26T18:42:00Z

@danmosemsft: If I find new issues or ideas I will definitely file an issue. But currently my free time is quite limited.

* Fix for Regex performance issue when compiled Regex is used with RegexOptions.IgnoreCase or RegexOptions.CultureInvariant. This saves over 40% in these cases. * Fix for Regex performance issue when compiled Regex is used with Rege…xOptions.IgnoreCase or RegexOptions.CultureInvariant. This saves over 40% in these cases. * Revert "Fix for Regex performance issue when compiled Regex is used with RegexOptions.IgnoreCase or RegexOptions.CultureInvariant." This reverts commit dotnet/corefx@e151e93. * Added TurkishI tests which check compiled and interpreted regular expressions. * Removed var of test Commit migrated from dotnet/corefx@58e2b4c

Alois-xx added 3 commits October 18, 2018 22:53

Fix for Regex performance issue when compiled Regex is used with Rege…

e151e93

…xOptions.IgnoreCase or RegexOptions.CultureInvariant. This saves over 40% in these cases.

Fix for Regex performance issue when compiled Regex is used with Rege…

62d7b7c

……xOptions.IgnoreCase or RegexOptions.CultureInvariant. This saves over 40% in these cases.

Revert "Fix for Regex performance issue when compiled Regex is used w…

07ce7fc

…ith RegexOptions.IgnoreCase or RegexOptions.CultureInvariant." This reverts commit e151e93.

danmoseley requested a review from ViktorHofer October 18, 2018 21:28

stephentoub reviewed Oct 18, 2018

View reviewed changes

ViktorHofer reviewed Oct 19, 2018

View reviewed changes

Added TurkishI tests which check compiled and interpreted regular exp…

a458fc0

…ressions.

stephentoub reviewed Oct 25, 2018

View reviewed changes

Removed var of test

2b18982

stephentoub merged commit 58e2b4c into dotnet:master Oct 25, 2018

Alois-xx deleted the Issue_32764 branch October 29, 2018 11:28

karelz added this to the 3.0 milestone Nov 15, 2018

Addresses Regex Perf Issue 32764 #32899

Addresses Regex Perf Issue 32764 #32899

Uh oh!

Conversation

Alois-xx commented Oct 18, 2018

Uh oh!

stephentoub Oct 18, 2018

Choose a reason for hiding this comment

Uh oh!

Alois-xx Oct 19, 2018

Choose a reason for hiding this comment

Uh oh!

stephentoub Oct 19, 2018

Choose a reason for hiding this comment

Uh oh!

stephentoub Oct 18, 2018

Choose a reason for hiding this comment

Uh oh!

ViktorHofer Oct 19, 2018

Choose a reason for hiding this comment

Uh oh!

stephentoub commented Oct 19, 2018

Uh oh!

ViktorHofer commented Oct 19, 2018

Uh oh!

ViktorHofer left a comment

Choose a reason for hiding this comment

Uh oh!

Alois-xx commented Oct 19, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Alois-xx commented Oct 20, 2018

Uh oh!

ViktorHofer commented Oct 22, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Alois-xx commented Oct 22, 2018

Uh oh!

Alois-xx commented Oct 24, 2018

Uh oh!

stephentoub Oct 25, 2018

Choose a reason for hiding this comment

Uh oh!

Alois-xx commented Oct 25, 2018

Uh oh!

stephentoub commented Oct 25, 2018

Uh oh!

danmoseley commented Oct 25, 2018

Uh oh!

Alois-xx commented Oct 26, 2018

Uh oh!

Uh oh!

Alois-xx commented Oct 19, 2018 •

edited

Loading

ViktorHofer commented Oct 22, 2018 •

edited

Loading