[BUG] Long line makes pattern matching (by //) hog Zsh

[BUG] Long line makes pattern matching (by //) hog Zsh

[Sebastian Gniazdowski]

[-- Attachment #1: Type: text/plain, Size: 568 bytes --]

Hello,
attached is a test program that in general blocks Zsh from ending.
Github user kapsh reported "fatal error: out of heap memory" and I
investigated that it's about long lines in his data. The pattern:

([^ /\\\\]##[^0-9/\\\\ ]##[^/\\\\]#(#e))

is:
1. not backslash nor slash nor space [^ /\\\\]##
2. not number, slash, backslash, space [^0-9/\\\\ ]##
3. not slash, backslash [^/\\\\]#
4. end of line (#e)

Removing first segment makes Zsh end after a couple of seconds (but
not on zsh-5.0.8). Tested on zsh-5.0.2, 5.0.8, 5.2

Best regards,
Sebastian Gniazdowski

[-- Attachment #2: 2pat.zsh --]
[-- Type: application/octet-stream, Size: 3284 bytes --]

#!/usr/bin/env zsh

emulate -LR zsh
setopt extendedglob

typeset -a list disp_list
list=( "3226  1802 /opt/google/chrome/chrome --type=renderer --enable-lcd-text --enable-features=DownloadResumption,IncidentReportingModuleLoadAnalysis<SafeBrowsingIncidentReportingServiceFeatures,IncidentReportingSuspiciousModuleReporting<SafeBrowsingIncidentReportingServiceFeatures,LinuxObsoleteSystemIsEndOfTheLine<LinuxObsoleteSystemIsEndOfTheLine,PrintPreviewDistiller,*UsePasswordSeparatedSigninFlow<PasswordSeparatedSigninFlow,WebFontsIntervention<WebFontsIntervention,*WebRTC-EnableWebRtcEcdsa<WebRTC-EnableWebRtcEcdsa,brotli-encoding<BrotliEncoding,enable-automatic-password-saving --disable-features=RenderingPipelineThrottling<RenderingPipelineThrottling --force-fieldtrials=AppBannerTriggering/Aggressive/*AutoReloadExperiment/FlagEnabled/*AutoReloadVisibleOnlyExperiment/FlagEnabled/AutofillProfileOrderByFrecency/Enabled/*BrotliEncoding/Enabled/CaptivePortalInterstitial/Enabled/*ChildAccountDetection/Disabled/*ClientSideDetectionModel/Model0/*CrossDevicePromo/28DaySingleProfile/EnableSessionCrashedBubbleUI/EnabledByFlag/*ExtensionActionRedesign/Enabled/*ExtensionDeveloperModeWarning/Default/*GFE/Default/InstanceID/Enabled/LinuxObsoleteSystemIsEndOfTheLine/EndOfLine/MaterialDesignDownloads/Enabled/*NewProfileManagement/Command-Line-Disabled/*OmniboxBundledExperimentV1/Unused_2/*OutOfProcessPac/Default/*PageRevisitInstrumentation/Default/PasswordBranding/Disabled/*PasswordGeneration/Disabled/*PasswordManagerSettingsMigration/Disable/PasswordSeparatedSigninFlow/Default/*QUIC/FlagEnabled/*RenderingPipelineThrottling/Disabled/ReportCertificateErrors/ShowAndPossiblySend/*ResourcePriorities/Launch50pct_11011_1_1_10/SHA1IdentityUIWarning/Enabled/SHA1ToolbarUIJanuary2016/Warning/SHA1ToolbarUIJanuary2017/Error/SSLCommonNameMismatchHandling/Enabled/*SafeBrowsingIncidentReportingService/Default/SafeBrowsingIncidentReportingServiceFeatures/WithSuspiciousModuleReporting/SafeBrowsingUnverifiedDownloads/DisableByParameterMostSbTypes2/SafeBrowsingUpdateFrequency/Default/*SimpleCacheTrial/ExperimentYes2/SyncHttpContentCompression/Disabled/*UMA-Population-Restrict/normal/*UMA-Uniformity-Trial-1-Percent/group_61/*UMA-Uniformity-Trial-10-Percent/default/*UMA-Uniformity-Trial-100-Percent/group_01/*UMA-Uniformity-Trial-20-Percent/group_03/*UMA-Uniformity-Trial-5-Percent/group_04/*UMA-Uniformity-Trial-50-Percent/group_01/*UseDelayAgnosticAEC/DefaultEnabled/*WebFontsIntervention/Enabled/WebRTC-EnableWebRtcEcdsa/Default/ --primordial-pipe-token=F02F1A91C1338A95D3C111314BA6B87A --lang=en-US --enable-crash-reporter=831AB4A2-FDF7-4D2B-9BE0-F22377348434, --extension-process --enable-webrtc-hw-h264-encoding --enable-offline-auto-reload --enable-offline-auto-reload-visible-only --show-saved-copy=secondary --blink-settings=fetchDeferLateScripts=true,fetchIncreaseFontPriority=true,fetchIncreasePriorities=true --enable-suggestions-with-substring-match -" )

_nlist_colorify_disp_list() {
    local col=$'\x1b[00;34m' reset=$'\x1b[0m'
    NLIST_COLORING_PATTERN="([^ /\\\\]##[^0-9/\\\\ ]##[^/\\\\]#(#e))"
    disp_list=( "${(@)disp_list//(#mi)$~NLIST_COLORING_PATTERN/$col${MATCH}$reset}" )
}

disp_list=( "${(@)list[1, 50]}" )
echo "Starting"
_nlist_colorify_disp_list
echo "Done"

Re: [BUG] Long line makes pattern matching (by //) hog Zsh

[Bart Schaefer]

On Jun 5,  4:36pm, Sebastian Gniazdowski wrote:
}
} 1. not backslash nor slash nor space [^ /\\\\]##
} 2. not number, slash, backslash, space [^0-9/\\\\ ]##
} 3. not slash, backslash [^/\\\\]#
} 4. end of line (#e)

It's in the block in pattern.c:patchmatch() that begins with the
comment:

		/*
		 * Lookahead to avoid useless matches. This is not possible
		 * with approximation.
		 */

Specifically, in the "if (no >= min) for (;;) ..." loop, at each charater
in the input string patmatch() is called recursively to look at the rest
of the string, which again enters this same loop because the next thing
is also a one-or-more expression, which calls recursively and again
enters the loop because the thing after that is a zero-or-more.

It consumes a LOT of memory while doing this, even if I add a hack to
prevent it from recursively re-entering lookahead (or to skip the
lookahead entirely).

Re: [BUG] Long line makes pattern matching (by //) hog Zsh

[Peter Stephenson]

0On Sun, 5 Jun 2016 12:10:20 -0700
Bart Schaefer <schaefer@brasslantern.com> wrote:

> On Jun 5,  4:36pm, Sebastian Gniazdowski wrote:
> }
> } 1. not backslash nor slash nor space [^ /\\\\]##
> } 2. not number, slash, backslash, space [^0-9/\\\\ ]##
> } 3. not slash, backslash [^/\\\\]#
> } 4. end of line (#e)
>
> It's in the block in pattern.c:patchmatch() that begins with the
> comment:
> 
> 		/*
> 		 * Lookahead to avoid useless matches. This is not possible
> 		 * with approximation.
> 		 */

That comment happens to be an irrelevance --- that's where it's looking
for something that's an exact match to follow.  There isn't one in this
case.  If there was, we would have latched onto it and we wouldn't need
to try quite so hard rearranging the other elements of the pattern.

The containing block is where it's handling # and ##, so there's no
great surprise it's'= there.

> Specifically, in the "if (no >= min) for (;;) ..." loop, at each charater
> in the input string patmatch() is called recursively to look at the rest
> of the string, which again enters this same loop because the next thing
> is also a one-or-more expression, which calls recursively and again
> enters the loop because the thing after that is a zero-or-more.

The problem is the patterns are pathological.  Each of them can match
the same characters.  So it's spending a lot of time repartitioning the
mathches  between the possibilities of 1. and 2. and 3. in the above.
That's not polynomially bounded.  I'm not sure if it's even
exponentially bounded.

What I'm not sure is if there's a way of improving this without some
special case or, obviously, making the patterns more specific.

pws

Re: [BUG] Long line makes pattern matching (by //) hog Zsh

[Peter Stephenson]

On Sun, 5 Jun 2016 20:37:08 +0100
Peter Stephenson <p.w.stephenson@ntlworld.com> wrote:
> The problem is the patterns are pathological.  Each of them can match
> the same characters.  So it's spending a lot of time repartitioning the
> mathches  between the possibilities of 1. and 2. and 3. in the above.
> That's not polynomially bounded.  I'm not sure if it's even
> exponentially bounded.
> 
> What I'm not sure is if there's a way of improving this without some
> special case or, obviously, making the patterns more specific.

We are maybe being naive in treating [...] expressions as if they are
"simple".  I think we have better protection against pathological
backtracking for more complicated patterns.

Does the following help?

pws

diff --git a/Src/pattern.c b/Src/pattern.c
index 4e2f236..bec980f 100644
--- a/Src/pattern.c
+++ b/Src/pattern.c
@@ -1425,7 +1425,7 @@ patcomppiece(int *flagp, int paren)
 	case Inbrack:
 	    DPUTS(zpc_special[ZPC_INBRACK] == Marker,
 		  "Treating '[' as pattern character although disabled");
-	    flags |= P_SIMPLE;
+	    /*flags |= P_SIMPLE;*/
 	    if (*patparse == Hat || *patparse == Bang) {
 		patparse++;
 		starter = patnode(P_ANYBUT);

Re: [BUG] Long line makes pattern matching (by //) hog Zsh

[Sebastian Gniazdowski]

On 5 June 2016 at 22:39, Peter Stephenson <p.w.stephenson@ntlworld.com> wrote:
> We are maybe being naive in treating [...] expressions as if they are
> "simple".  I think we have better protection against pathological
> backtracking for more complicated patterns.
>
> Does the following help?

It's slower with this patch:

3,02s user 0,02s system 99% cpu 3,064 total
vs
1,55s user 0,24s system 93% cpu 1,909 total

for pattern with first segment removed, i.e.: ([^0-9/\\\\ ]##[^/\\\\]#(#e))

Best regards,
Sebastian Gniazdowski

Re: [BUG] Long line makes pattern matching (by //) hog Zsh

[Bart Schaefer]

On Jun 5,  8:37pm, Peter Stephenson wrote:
}
} > On Jun 5,  4:36pm, Sebastian Gniazdowski wrote:
} > }
} > } 1. not backslash nor slash nor space [^ /\\\\]##
} > } 2. not number, slash, backslash, space [^0-9/\\\\ ]##
} > } 3. not slash, backslash [^/\\\\]#
} > } 4. end of line (#e)
} 
} The problem is the patterns are pathological.  Each of them can match
} the same characters.

Indeed, given that the first sub-pattern is "at least one" and the
second sub-pattern includes all the same characters as the first, I
think the first ## can be removed without changing what is matched;
i.e.:

    NLIST_COLORING_PATTERN="([^ /\\\\][^0-9/\\\\ ]##[^/\\\\]#(#e))"

And with that pattern the whole test script completes in a couple of
seconds.

} So it's spending a lot of time repartitioning the
} mathches  between the possibilities of 1. and 2. and 3. in the above.
} That's not polynomially bounded.  I'm not sure if it's even
} exponentially bounded.

I tried to work that out and came up with O(n**((n-1)**(n-2))) before
deciding I didn't remember enough of that math to believe it.

The other problem, which I hadn't really thought about before, is that
this is all being applied inside ${str//pat/repl} expression, bounded
only on the right, which means it has to try every position in $str
even when the pattern does NOT match, which for a fair chunk of the
example input it does not.

There may be an optimization possible in the way that igetmatch()
interacts with pattrylen(), so as to skip forward over the longest
NON-match as well as skipping over the longest match.

Re: [BUG] Long line makes pattern matching (by //) hog Zsh

[Sebastian Gniazdowski]

On 5 June 2016 at 21:10, Bart Schaefer <schaefer@brasslantern.com> wrote:
> It's in the block in pattern.c:patchmatch() that begins with the
> comment:
>
>                 /*
>                  * Lookahead to avoid useless matches. This is not possible
>                  * with approximation.
>                  */
>
> Specifically, in the "if (no >= min) for (;;) ..." loop, at each charater
> in the input string patmatch() is called recursively to look at the rest
> of the string, which again enters this same loop because the next thing
> is also a one-or-more expression, which calls recursively and again
> enters the loop because the thing after that is a zero-or-more.
>
> It consumes a LOT of memory while doing this, even if I add a hack to
> prevent it from recursively re-entering lookahead (or to skip the
> lookahead entirely).

Too bad the input isn't even that large, ~2850 characters. And that
the pattern even with first segment removed – ([^0-9/\\\\
]##[^/\\\\]#(#e)) – i.e. two 0/1-or-more segments, is also slow,
although doesn't hog Zsh. Maybe your changes would make a difference
for that input size, and for first segment removed, i.e. two # or ##
blocks? I had simplified highlighting in my tool and should be fine
now, but maybe in future I could restore the more sophisticated
pattern.

Best regards,
Sebastian Gniazdowski